Merge branch 'develop' into feature/increase_cpu

benchmark/paddle/image/vgg.py

@@ -13,7 +13,7 @@ define_py_data_sources2(
 
 settings(
     batch_size=batch_size,
-    learning_rate=0.01 / batch_size,
+    learning_rate=0.001 / batch_size,
     learning_method=MomentumOptimizer(0.9),
     regularization=L2Regularization(0.0005 * batch_size))
 

cmake/cblas.cmake

 # Find the CBlas and lapack libraries
 #
-# It will search MKL, atlas, OpenBlas, reference-cblas in order.
+# It will search MKLML, atlas, OpenBlas, reference-cblas in order.
 #
 # If any cblas implementation found, the following variable will be set.
-#    CBLAS_PROVIDER  # one of MKL, ATLAS, OPENBLAS, REFERENCE
+#    CBLAS_PROVIDER  # one of MKLML, ATLAS, OPENBLAS, REFERENCE
 #    CBLAS_INC_DIR   # the include directory for cblas.
 #    CBLAS_LIBS      # a list of libraries should be linked by paddle.
 #                    # Each library should be full path to object file.
-#
-# User should set one of MKL_ROOT, ATLAS_ROOT, OPENBLAS_ROOT, REFERENCE_CBLAS_ROOT
-# during cmake. If none of them set, it will try to find cblas implementation in
-# system paths.
-#
 
 set(CBLAS_FOUND OFF)
 
@@ -30,44 +25,6 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
   return()
 endif()
 
-## Then find MKL.
-set(INTEL_MKL_ROOT "/opt/intel/mkl" CACHE PATH "Folder contains intel mkl libs")
-set(MKL_ROOT $ENV{MKL_ROOT} CACHE PATH "Folder contains env MKL")
-
-set(MKL_INCLUDE_SEARCH_PATHS
-  ${MKL_ROOT}/include
-  ${INTEL_MKL_ROOT}/include)
-set(MKL_LIB_SEARCH_PATHS
-  ${MKL_ROOT}/lib
-  ${MKL_ROOT}/lib/intel64
-  ${INTEL_MKL_ROOT}/lib
-  ${INTEL_MKL_ROOT}/lib/intel64)
-
-find_path(MKL_INC_DIR mkl.h PATHS
-  ${MKL_INCLUDE_SEARCH_PATHS})
-find_path(MKL_LAPACK_INC_DIR mkl_lapacke.h PATHS
-  ${MKL_INCLUDE_SEARCH_PATHS})
-find_library(MKL_CORE_LIB NAMES mkl_core PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-find_library(MKL_SEQUENTIAL_LIB NAMES mkl_sequential PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-find_library(MKL_INTEL_LP64 NAMES mkl_intel_lp64 PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-
-if(MKL_LAPACK_INC_DIR AND MKL_INC_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
-  set(CBLAS_FOUND ON)
-  set(CBLAS_PROVIDER MKL)
-  set(CBLAS_INC_DIR ${MKL_INC_DIR} ${MKL_LAPACK_INC_DIR})
-  set(CBLAS_LIBRARIES ${MKL_INTEL_LP64} ${MKL_SEQUENTIAL_LIB} ${MKL_CORE_LIB})
-
-  add_definitions(-DPADDLE_USE_MKL)
-  add_definitions(-DLAPACK_FOUND)
-
-  message(STATUS "Found MKL (include: ${MKL_INC_DIR}, library: ${CBLAS_LIBRARIES})")
-  message(STATUS "Found lapack in MKL (include: ${MKL_LAPACK_INC_DIR})")
-  return()
-endif()
-
 ## Then find atlas.
 set(ATLAS_ROOT $ENV{ATLAS_ROOT} CACHE PATH "Folder contains Atlas")
 set(ATLAS_INCLUDE_SEARCH_PATHS

cmake/external/mkldnn.cmake

@@ -46,16 +46,20 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
 ENDIF()
 
+SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
+SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow")
 ExternalProject_Add(
     ${MKLDNN_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "v0.10"
+    GIT_TAG             "v0.11"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
     CMAKE_ARGS          -DMKLROOT=${MKLDNN_MKLROOT}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
                         -DMKLROOT:PATH=${MKLDNN_MKLROOT}
 )

cmake/external/mklml.cmake

@@ -27,8 +27,8 @@ ENDIF()
 INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
-SET(MKLML_VER           "mklml_lnx_2018.0.20170720")
-SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz")
+SET(MKLML_VER           "mklml_lnx_2018.0.1.20171007")
+SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.11/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")

cmake/external/openblas.cmake

@@ -115,7 +115,7 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
 SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
 FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
-IF(${CBLAS_PROVIDER} MATCHES MKL)
+IF(${CBLAS_PROVIDER} EQUAL MKLML)
     ADD_LIBRARY(cblas SHARED ${dummyfile})
 ELSE()
     ADD_LIBRARY(cblas STATIC ${dummyfile})

doc/api/v2/data.rst

@@ -2,112 +2,9 @@
 Data Reader Interface and DataSets
 ==================================
 
+..  toctree::
+    :maxdepth: 1
 
-DataTypes
-=========
-
-..  automodule:: paddle.v2.data_type
-    :members:
-    :noindex:
-
-DataFeeder
-==========
-
-..  automodule:: paddle.v2.data_feeder
-    :members:
-    :noindex:
-
-Reader
-======
-
-..  automodule:: paddle.v2.reader
-    :members:
-    :noindex:
-
-..  automodule:: paddle.v2.reader.creator
-    :members:
-    :noindex:
-
-minibatch
-=========
-
-..  automodule:: paddle.v2.minibatch
-    :members:
-    :noindex:
-
-Dataset
-=======
-
-..  automodule:: paddle.v2.dataset
-    :members:
-    :noindex:
-
-mnist
-+++++
-
-..  automodule:: paddle.v2.dataset.mnist
-    :members:
-    :noindex:
-
-cifar
-+++++
-
-..  automodule:: paddle.v2.dataset.cifar
-    :members:
-    :noindex:
-
-conll05
-+++++++
-
-..  automodule:: paddle.v2.dataset.conll05
-    :members: get_dict,get_embedding,test
-    :noindex:
-
-imdb
-++++
-
-..  automodule:: paddle.v2.dataset.imdb
-    :members:
-    :noindex:
-
-imikolov
-++++++++
-
-..  automodule:: paddle.v2.dataset.imikolov
-    :members:
-    :noindex:
-
-movielens
-+++++++++
-
-..  automodule:: paddle.v2.dataset.movielens
-    :members:
-    :noindex:
-
-..  autoclass:: paddle.v2.dataset.movielens.MovieInfo
-    :noindex:
-    
-..  autoclass:: paddle.v2.dataset.movielens.UserInfo
-    :noindex:
-
-sentiment
-+++++++++
-
-..  automodule:: paddle.v2.dataset.sentiment
-    :members:
-    :noindex:
-
-uci_housing
-+++++++++++
-
-..  automodule:: paddle.v2.dataset.uci_housing
-    :members:
-    :noindex:
-
-wmt14
-+++++
-
-..  automodule:: paddle.v2.dataset.wmt14
-    :members:
-    :noindex:
-
+    data/data_reader.rst
+    data/image.rst
+    data/dataset.rst

doc/api/v2/data/data_reader.rst

+=====================
+Data Reader Interface
+=====================
+
+
+DataTypes
+=========
+
+..  automodule:: paddle.v2.data_type
+    :members:
+    :noindex:
+
+DataFeeder
+==========
+
+..  automodule:: paddle.v2.data_feeder
+    :members:
+    :noindex:
+
+Reader
+======
+
+..  automodule:: paddle.v2.reader
+    :members:
+    :noindex:
+
+..  automodule:: paddle.v2.reader.creator
+    :members:
+    :noindex:
+
+minibatch
+=========
+
+..  automodule:: paddle.v2.minibatch
+    :members:
+    :noindex:

doc/api/v2/data/dataset.rst

+Dataset
+=======
+
+..  automodule:: paddle.v2.dataset
+    :members:
+    :noindex:
+
+mnist
++++++
+
+..  automodule:: paddle.v2.dataset.mnist
+    :members:
+    :noindex:
+
+cifar
++++++
+
+..  automodule:: paddle.v2.dataset.cifar
+    :members:
+    :noindex:
+
+conll05
++++++++
+
+..  automodule:: paddle.v2.dataset.conll05
+    :members: get_dict,get_embedding,test
+    :noindex:
+
+imdb
+++++
+
+..  automodule:: paddle.v2.dataset.imdb
+    :members:
+    :noindex:
+
+imikolov
+++++++++
+
+..  automodule:: paddle.v2.dataset.imikolov
+    :members:
+    :noindex:
+
+movielens
++++++++++
+
+..  automodule:: paddle.v2.dataset.movielens
+    :members:
+    :noindex:
+
+..  autoclass:: paddle.v2.dataset.movielens.MovieInfo
+    :noindex:
+    
+..  autoclass:: paddle.v2.dataset.movielens.UserInfo
+    :noindex:
+
+sentiment
++++++++++
+
+..  automodule:: paddle.v2.dataset.sentiment
+    :members:
+    :noindex:
+
+uci_housing
++++++++++++
+
+..  automodule:: paddle.v2.dataset.uci_housing
+    :members:
+    :noindex:
+
+wmt14
++++++
+
+..  automodule:: paddle.v2.dataset.wmt14
+    :members:
+    :noindex:

doc/api/v2/data/image.rst

+Image Interface
+===============
+
+..  automodule:: paddle.v2.image
+    :members:

doc/design/ops/sequence_decoder.md

+# Design: Sequence Decoder Generating LoDTensors
+In tasks such as machine translation and image to text, 
+a [sequence decoder](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.md) is necessary to generate sequences.
+
+This documentation describes how to implement the sequence decoder as an operator.
+
+## Beam Search based Decoder
+The [beam search algorithm](https://en.wikipedia.org/wiki/Beam_search) is necessary when generating sequences, 
+it is a heuristic search algorithm that explores the paths by expanding the most promising node in a limited set.
+
+In the old version of PaddlePaddle, a C++ class `RecurrentGradientMachine` implements the general sequence decoder based on beam search, 
+due to the complexity, the implementation relays on a lot of special data structures, 
+quite trivial and hard to be customized by users.
+
+There are a lot of heuristic tricks in the sequence generation tasks, 
+so the flexibility of sequence decoder is very important to users.
+
+During PaddlePaddle's refactoring work,
+some new concept is proposed such as [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) and [TensorArray](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/tensor_array.md) that can better support sequence usage,
+and they can help to make the implementation of beam search based sequence decoder **more transparent and modular** .
+
+For example, the RNN sates, candidates IDs and probabilities of beam search can be represented as `LoDTensors`;
+the selected candidate's IDs in each time step can be stored in a `TensorArray`, and `Packed` to the sentences translated.
+
+## Changing LoD's absolute offset to relative offsets
+The current `LoDTensor` is designed to store levels of variable-length sequences,
+it stores several arrays of integers each represents a level.
+
+The integers in each level represents the begin and end (not inclusive) offset of a sequence **in the underlying tensor**, 
+let's call this format the **absolute-offset LoD** for clear.
+
+The relative-offset LoD can fast retrieve any sequence but fails to represent empty sequences, for example, a two-level LoD is as follows
+```python
+[[0, 3, 9]
+ [0, 2, 3, 3, 3, 9]]
+```
+The first level tells that there are two sequences:
+- the first's offset is `[0, 3)`
+- the second's offset is `[3, 9)`
+
+while on the second level, there are several empty sequences that both begin and end at `3`.
+It is impossible to tell how many empty second-level sequences exist in the first-level sequences.
+
+There are many scenarios that relay on empty sequence representation,
+such as machine translation or image to text, one instance has no translations or the empty candidate set for a prefix.
+
+So let's introduce another format of LoD, 
+it stores **the offsets of the lower level sequences** and is called **relative-offset** LoD.
+
+For example, to represent the same sequences of the above data
+
+```python
+[[0, 3, 6]
+ [0, 2, 3, 3, 3, 9]]
+```
+
+the first level represents that there are two sequences, 
+their offsets in the second-level LoD is `[0, 3)` and `[3, 5)`.
+
+The second level is the same with the relative offset example because the lower level is a tensor.
+It is easy to find out the second sequence in the first-level LoD has two empty sequences.
+
+The following demos are based on relative-offset LoD.
+
+## Usage in a simple machine translation model
+Let's start from a simple machine translation model that is simplified from [machine translation chapter](https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation) to draw a simple blueprint of what a sequence decoder can do and how to use it.
+
+The model has an encoder that learns the semantic vector from a sequence,
+and a decoder which uses the sequence decoder to generate new sentences.
+
+**Encoder**
+```python
+import paddle as pd
+
+dict_size = 8000
+source_dict_size = dict_size
+target_dict_size = dict_size
+word_vector_dim = 128
+encoder_dim = 128
+decoder_dim = 128
+beam_size = 5
+max_length = 120
+
+# encoder
+src_word_id = pd.data(
+    name='source_language_word',
+    type=pd.data.integer_value_sequence(source_dict_dim))
+src_embedding = pd.embedding(size=source_dict_size, size=word_vector_dim)
+
+src_word_vec = pd.lookup(src_embedding, src_word_id)
+
+encoder_out_seq = pd.gru(input=src_word_vec, size=encoder_dim)
+
+encoder_ctx = pd.last_seq(encoder_out_seq)
+# encoder_ctx_proj is the learned semantic vector
+encoder_ctx_proj = pd.fc(
+    encoder_ctx, size=decoder_dim, act=pd.activation.Tanh(), bias=None)
+```
+
+**Decoder**
+
+```python
+def generate():
+    decoder = pd.while_loop()
+    with decoder.step():
+        decoder_mem = decoder.memory(init=encoder_ctx)  # mark the memory
+        generated_ids = decoder.memory() # TODO init to batch_size <s>s
+        generated_scores = decoder.memory() # TODO init to batch_size 1s or 0s
+
+        target_word = pd.lookup(trg_embedding, gendrated_ids)
+        # expand encoder_ctx's batch to fit target_word's lod
+        # for example
+        # decoder_mem.lod is
+        # [[0 1 3],
+        #  [0 1 3 6]]
+        # its tensor content is [a1 a2 a3 a4 a5]
+        # which means there are 2 sentences to translate
+        #   - the first sentence has 1 translation prefixes, the offsets are [0, 1)
+        #   - the second sentence has 2 translation prefixes, the offsets are [1, 3) and [3, 6)
+        # the target_word.lod is 
+        # [[0, 1, 6]
+        #  [0, 2, 4, 7, 9 12]]
+        # which means 2 sentences to translate, each has 1 and 5 prefixes
+        # the first prefix has 2 candidates
+        # the following has 2, 3, 2, 3 candidates
+        # the encoder_ctx_expanded's content will be
+        # [a1 a1 a2 a2 a3 a3 a3 a4 a4 a5 a5 a5]
+        encoder_ctx_expanded = pd.lod_expand(encoder_ctx, target_word)
+        decoder_input = pd.fc(
+            act=pd.activation.Linear(),
+            input=[target_word, encoder_ctx],
+            size=3 * decoder_dim)
+        gru_out, cur_mem = pd.gru_step(
+            decoder_input, mem=decoder_mem, size=decoder_dim)
+        scores = pd.fc(
+            gru_out,
+            size=trg_dic_size,
+            bias=None,
+            act=pd.activation.Softmax())
+        # K is an config
+        topk_scores, topk_ids = pd.top_k(scores, K)
+        topk_generated_scores = pd.add_scalar(topk_scores, generated_scores)
+
+        selected_ids, selected_generation_scores = decoder.beam_search(
+            topk_ids, topk_generated_scores)
+
+        # update the states
+        decoder_mem.update(cur_mem)  # tells how to update state
+        generated_ids.update(selected_ids)
+        generated_scores.update(selected_generation_scores)
+
+        decoder.output(selected_ids)
+        decoder.output(selected_generation_scores)
+
+translation_ids, translation_scores = decoder()
+```
+The `decoder.beam_search` is a operator that given the candidates and the scores of translations including the candidates,
+return the result of the beam search algorithm.
+
+In this way, users can customize anything on the inputs or outputs of beam search, for example, two ways to prune some translation prefixes
+
+1. meke the correspondind elements in `topk_generated_scores` zero or some small values, beam_search will discard this candidate.
+2. remove some specific candidate in `selected_ids`
+3. get the final `translation_ids`, remove the translation sequence in it.
+
+The implementation of sequence decoder can reuse the C++ class [RNNAlgorithm](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/paddle/operators/dynamic_recurrent_op.h#L30),
+so the python syntax is quite similar to a [RNN](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/doc/design/block.md#blocks-with-for-and-rnnop).
+
+Both of them are two-level `LoDTensors`
+
+- the first level represents `batch_size` of (source) sentences;
+- the second level represents the candidate ID sets for translation prefix.
+
+for example, 3 source sentences to translate, and has 2, 3, 1 candidates.
+
+Unlike an RNN, in sequence decoder, the previous state and the current state have different LoD and shape,
+a `lod_expand` operator is used to expand the LoD of the previous state to fit the current state.
+
+For example, the previous state
+
+* LoD is `[0, 1, 3][0, 2, 5, 6]`
+* content of tensor is `a1 a2 b1 b2 b3 c1`
+
+the current state stored in `encoder_ctx_expanded`
+
+* LoD is `[0, 2, 7][0 3 5 8 9 11 11]`
+* the content is 
+  - a1 a1 a1 (a1 has 3 candidates, so the state should be copied 3 times for each candidates)
+  - a2 a2
+  - b1 b1 b1
+  - b2
+  - b3 b3
+  - None (c1 has 0 candidates, so c1 is dropped)
+
+Benefit from the relative offset LoD, empty candidate set can be represented naturally.
+
+the status in each time step can be stored in `TensorArray`, and `Pack`ed to a final LoDTensor, the corresponding syntax is 
+
+```python
+decoder.output(selected_ids)
+decoder.output(selected_generation_scores)
+```
+
+the `selected_ids` is the candidate ids for the prefixes, 
+it will be `Packed` by `TensorArray` to a two-level `LoDTensor`,
+the first level represents the source sequences,
+the second level represents generated sequences.
+
+Pack the `selected_scores` will get a `LoDTensor` that stores scores of each candidate of translations.
+
+Pack the `selected_generation_scores` will get a `LoDTensor`, and each tail is the probability of the translation.
+
+## LoD and shape changes during decoding
+<p align="center">
+  <img src="./images/LOD-and-shape-changes-during-decoding.jpg"/>
+</p>
+
+According the image above, the only phrase to change LoD is beam search.
+
+## Beam search design
+The beam search algorthm will be implemented as one method of the sequence decoder, it has 3 inputs
+
+1. `topk_ids`, top K candidate ids for each prefix.
+2. `topk_scores`, the corresponding scores for `topk_ids`
+3. `generated_scores`, the score of the prefixes.
+
+All of the are LoDTensors, so that the sequence affilication is clear.
+Beam search will keep a beam for each prefix and select a smaller candidate set for each prefix.
+
+It will return three variables
+
+1. `selected_ids`, the final candidate beam search function selected for the next step.
+2. `selected_scores`, the scores for the candidates.
+3. `generated_scores`, the updated scores for each prefixes (with the new candidates appended).
+
+## Introducing the LoD-based `Pack` and `Unpack` methods in `TensorArray`
+The `selected_ids`, `selected_scores` and `generated_scores` are LoDTensors,
+and they exist in each time step,
+so it is natural to store them in arrays.
+
+Currently, PaddlePaddle has a module called `TensorArray` which can store an array of tensors,
+the results of beam search are better to store in a `TensorArray`.
+
+The `Pack` and `UnPack` in `TensorArray` are used to package tensors in the array to a `LoDTensor` or split the `LoDTensor` to an array of tensors. 
+It needs some extensions to support pack or unpack an array of `LoDTensors`.

paddle/framework/backward_test.cc

@@ -21,7 +21,7 @@
 #include "paddle/framework/var_desc.h"
 #include "paddle/operators/net_op.h"
 
-USE_OP(fill_constant);
+USE_NO_KERNEL_OP(fill_constant);
 
 namespace paddle {
 namespace framework {

paddle/framework/data_type.h

@@ -34,6 +34,21 @@ inline DataType ToDataType(std::type_index type) {
   }
 }
 
+inline std::type_index ToTypeIndex(DataType type) {
+  switch (type) {
+    case DataType::FP32:
+      return typeid(float);
+    case DataType::FP64:
+      return typeid(double);
+    case DataType::INT32:
+      return typeid(int);
+    case DataType::INT64:
+      return typeid(int64_t);
+    default:
+      PADDLE_THROW("Not support type %d", type);
+  }
+}
+
 template <typename Visitor>
 inline void VisitDataType(DataType type, Visitor visitor) {
   switch (type) {

paddle/framework/ddim.cc

@@ -79,6 +79,13 @@ DDim make_ddim(const std::vector<int64_t>& dims) {
   return result;
 }
 
+DDim make_ddim(const std::vector<int>& dims) {
+  std::vector<int64_t> res(dims.size());
+  std::transform(dims.begin(), dims.end(), res.begin(),
+                 [](int d) { return static_cast<int64_t>(d); });
+  return make_ddim(res);
+}
+
 /// @cond HIDDEN
 // XXX For some reason, putting this in an anonymous namespace causes errors
 class DynamicMutableIndexer : public boost::static_visitor<int64_t&> {

paddle/framework/ddim.h

@@ -81,6 +81,8 @@ struct DDim {
  */
 DDim make_ddim(const std::vector<int64_t>& dims);
 
+DDim make_ddim(const std::vector<int>& dims);
+
 /**
  * \brief Make a DDim from an initializer list
  *

paddle/function/CMakeLists.txt

@@ -45,6 +45,7 @@ if(WITH_GPU)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
     add_simple_unittest(SwitchOpTest)
+    add_simple_unittest(ScaleSubRegionOpTest)
 endif()
 
 add_simple_unittest(Im2ColTest)

paddle/function/FunctionTest.h

@@ -110,6 +110,7 @@ public:
         function2_(FunctionBase::funcRegistrar_.createByType(name2)) {
     function1_->init(config);
     function2_->init(config);
+    initArgsCallback_ = nullptr;
   }
 
   ~Compare2Function() {}
@@ -170,6 +171,10 @@ public:
                                       *seq2_));
   }
 
+  void registerInitCallback(std::function<void(BufferArg&, size_t)> callback) {
+    initArgsCallback_ = callback;
+  }
+
   // output need only contains shape, do not contains data.
   void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) {
     size_t size =
@@ -340,6 +345,10 @@ protected:
         initArg(*func1Inputs_[i]);
       }
 
+      if (initArgsCallback_ != nullptr) {
+        initArgsCallback_(*func1Inputs_[i], i);
+      }
+
       copyArg_(*func1Inputs_[i], *func2Inputs_[i]);
     }
   }
@@ -386,6 +395,7 @@ protected:
   std::shared_ptr<SequenceIdArg> seq1_;
   std::shared_ptr<SequenceIdArg> seq2_;
   test::CopyArgument<DType1, DType2> copyArg_;
+  std::function<void(BufferArg&, size_t)> initArgsCallback_;
 };
 
 class CpuGpuFuncCompare

paddle/function/ScaleSubRegionOp.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "ScaleSubRegionOp.h"
+#include "paddle/function/TensorShape.h"
+
+namespace paddle {
+
+template <>
+void ScaleSubRegion<DEVICE_TYPE_CPU>(real* outputs,
+                                     const real* inputs,
+                                     const real* indices,
+                                     const TensorShape shape,
+                                     const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  memcpy(outputs, inputs, number * channel * height * width * sizeof(real));
+
+  for (int n = 0; n < number; ++n) {
+    // indices start from 1
+    int offset = n * 6;
+    for (int c = indices[offset] - 1; c < indices[offset + 1]; ++c) {
+      for (int h = indices[offset + 2] - 1; h < indices[offset + 3]; ++h) {
+        for (int w = indices[offset + 4] - 1; w < indices[offset + 5]; ++w) {
+          int idx = ((n * channel + c) * height + h) * width + w;
+          outputs[idx] *= value;
+        }
+      }
+    }
+  }
+}
+
+template <>
+void ScaleSubRegionGrad<DEVICE_TYPE_CPU>(const real* inGrad,
+                                         real* outGrad,
+                                         const real* indices,
+                                         const TensorShape shape,
+                                         const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  for (int n = 0; n < number; ++n) {
+    for (int c = 0; c < channel; ++c) {
+      for (int h = 0; h < height; ++h) {
+        for (int w = 0; w < width; ++w) {
+          int idx = ((n * channel + c) * height + h) * width + w;
+          int offset = n * 6;
+          if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+              h >= (indices[offset + 2] - 1) &&
+              h <= (indices[offset + 3] - 1) &&
+              w >= (indices[offset + 4] - 1) &&
+              w <= (indices[offset + 5] - 1)) {
+            outGrad[idx] += inGrad[idx] * value;
+          } else {
+            outGrad[idx] += inGrad[idx];
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief For each instance, ScaleSubRegion can be used to multiply a value to
+ *        a specified sub continuous region. By providing start index and end
+ *        index for C/H/W, you can specify the location and shape of the region.
+ *
+ * Argument in this Function:
+ * \param inputs    A 4-D tensor with shape [N, C, H, W], only one input.
+ * \param indices   A 2-D tensor with shape [N, 6], indicates the sub region.
+ * \param outputs   A 4-D tensor with same shape as inputs, output value.
+ */
+template <DeviceType Device>
+class ScaleSubRegionFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(2UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
+
+    TensorShape shape = inputs[0].shape();
+
+    ScaleSubRegion<Device>(outputs[0].data<real>(),
+                           inputs[0].data<real>(),
+                           inputs[1].data<real>(),
+                           shape,
+                           conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+/**
+ * \brief The backward propagation of ScaleSubRegion Function.
+ *
+ * Argument in this Function:
+ * \param inputs  A 4-D tensor with shape [N, C, H, W], output gradient.
+ * \param indices A 2-D tensor with shape [N, 6], indicates the sub region.
+ * \param outputs A 4-D tensor with shape [N, C, H, W], gradient of input value.
+ */
+
+template <DeviceType Device>
+class ScaleSubRegionGradFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(2UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
+
+    TensorShape shape = inputs[0].shape();
+
+    ScaleSubRegionGrad<Device>(inputs[0].data<real>(),
+                               outputs[0].data<real>(),
+                               inputs[1].data<real>(),
+                               shape,
+                               conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+REGISTER_TYPED_FUNC(ScaleSubRegion, CPU, ScaleSubRegionFunc);
+REGISTER_TYPED_FUNC(ScaleSubRegionGrad, CPU, ScaleSubRegionGradFunc);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_TYPED_FUNC(ScaleSubRegion, GPU, ScaleSubRegionFunc);
+REGISTER_TYPED_FUNC(ScaleSubRegionGrad, GPU, ScaleSubRegionGradFunc);
+#endif
+
+}  // namespace paddle

paddle/function/ScaleSubRegionOp.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief Function to multiply a value to values in specified sub continuous
+ *        region. Indices must be provided to indcate the location and shape of
+ *        the region and the multiplied value is passed by configure variable.
+ *
+ *
+ * \param[out] outputs  Output value.
+ * \param[in]  inputs   Input data which contains NCHW information.
+ * \param[in]  indices  Indices data to indcate the sub region.
+ * \param[in]  shape    Tensor shape of input value.
+ * \param[in]  conf     Configure variable which contains the multiplied value.
+ */
+template <DeviceType Device>
+void ScaleSubRegion(real* outputs,
+                    const real* inputs,
+                    const real* indices,
+                    const TensorShape shape,
+                    const FuncConfig& conf);
+
+/**
+ * \brief Backward propagation function of ScaleSubRegion.
+ *
+ * \param[out] inGrad   Gradients of previous layer.
+ * \param[in]  outGrad  Output gradient.
+ * \param[in]  indices  Indices data.
+ * \param[in]  shape    The Shape of input tensor.
+ * \param[in]  conf     Configure variable.
+ */
+template <DeviceType Device>
+void ScaleSubRegionGrad(const real* inGrad,
+                        real* outGrad,
+                        const real* indices,
+                        const TensorShape shape,
+                        const FuncConfig& conf);
+}  // namespace paddle

paddle/function/ScaleSubRegionOpGpu.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "ScaleSubRegionOp.h"
+#include "hl_base.h"
+
+namespace paddle {
+
+__global__ void KeScaleSubRegion(real* outputs,
+                                 const real* inputs,
+                                 const real* indices,
+                                 real value,
+                                 int channel,
+                                 int height,
+                                 int width,
+                                 int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % width;
+    const int h = (idx / width) % height;
+    const int c = (idx / width / height) % channel;
+    const int n = idx / width / height / channel;
+
+    const int offset = n * 6;
+    if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+        h >= (indices[offset + 2] - 1) && h <= (indices[offset + 3] - 1) &&
+        w >= (indices[offset + 4] - 1) && w <= (indices[offset + 5] - 1)) {
+      outputs[idx] = inputs[idx] * value;
+    } else {
+      outputs[idx] = inputs[idx];
+    }
+  }
+}
+
+template <>
+void ScaleSubRegion<DEVICE_TYPE_GPU>(real* outputs,
+                                     const real* inputs,
+                                     const real* indices,
+                                     const TensorShape shape,
+                                     const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  size_t nth = number * channel * height * width;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeScaleSubRegion<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      outputs, inputs, indices, value, channel, height, width, nth);
+  CHECK_SYNC("ScaleSubRegion");
+}
+
+__global__ void KeScaleSubRegionDiff(const real* inGrad,
+                                     real* outGrad,
+                                     const real* indices,
+                                     real value,
+                                     int channel,
+                                     int height,
+                                     int width,
+                                     int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % width;
+    const int h = (idx / width) % height;
+    const int c = (idx / width / height) % channel;
+    const int n = idx / width / height / channel;
+
+    const int offset = n * 6;
+    if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+        h >= (indices[offset + 2] - 1) && h <= (indices[offset + 3] - 1) &&
+        w >= (indices[offset + 4] - 1) && w <= (indices[offset + 5] - 1)) {
+      outGrad[idx] += inGrad[idx] * value;
+    } else {
+      outGrad[idx] += inGrad[idx];
+    }
+  }
+}
+
+template <>
+void ScaleSubRegionGrad<DEVICE_TYPE_GPU>(const real* inGrad,
+                                         real* outGrad,
+                                         const real* indices,
+                                         const TensorShape shape,
+                                         const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  size_t nth = number * channel * height * width;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeScaleSubRegionDiff<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      inGrad, outGrad, indices, value, channel, height, width, nth);
+  CHECK_SYNC("ScaleSubRegionGrad");
+}
+
+}  // namespace paddle

paddle/function/ScaleSubRegionOpTest.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "FunctionTest.h"
+
+namespace paddle {
+
+TEST(ScaleSubRegion, real) {
+  for (size_t numSamples : {5, 32}) {
+    for (size_t channels : {5, 32}) {
+      for (size_t imgSizeH : {5, 33}) {
+        for (size_t imgSizeW : {5, 32}) {
+          for (real value : {-0.5, 0.0, 0.5}) {
+            for (bool firstHalf : {false, true}) {
+              VLOG(3) << " numSamples=" << numSamples
+                      << " channels=" << channels << " imgSizeH=" << imgSizeH
+                      << " imgSizeW=" << imgSizeW;
+
+              for (bool testGrad : {false, true}) {
+                CpuGpuFuncCompare compare(
+                    testGrad ? "ScaleSubRegionGrad" : "ScaleSubRegion",
+                    FuncConfig().set<real>("value", value));
+
+                TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
+                TensorShape indicesShape{numSamples, 6};
+
+                compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+                compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, indicesShape));
+
+                compare.registerInitCallback([=](BufferArg& arg, size_t index) {
+                  if (index == 1) {
+                    real* data = (real*)arg.data();
+
+                    for (size_t i = 0; i < numSamples; ++i) {
+                      size_t offset = i * 6;
+                      data[offset] = firstHalf ? 1 : channels / 2;
+                      data[offset + 1] = firstHalf ? channels / 2 : channels;
+                      data[offset + 2] = firstHalf ? 1 : imgSizeH / 2;
+                      data[offset + 3] = firstHalf ? imgSizeH / 2 : imgSizeH;
+                      data[offset + 4] = firstHalf ? 1 : imgSizeW / 2;
+                      data[offset + 5] = firstHalf ? imgSizeW / 2 : imgSizeW;
+                    }
+                  }
+                });
+
+                compare.addOutputs(
+                    BufferArg(
+                        VALUE_TYPE_FLOAT, shape, testGrad ? ADD_TO : ASSIGN_TO),
+                    testGrad ? ADD_TO : ASSIGN_TO);
+                compare.run();
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNAddtoLayer.cpp

@@ -62,16 +62,14 @@ void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,
                                 MKLDNNMatrixPtr& wgt,
                                 MKLDNNMatrixPtr& bias,
                                 MKLDNNMatrixPtr& out) {
-  if (biases_) {
-    LOG(FATAL) << "not implemented yet";
-  }
-  resetFwdBuffers(inVals_, out);
+  resetFwdBuffers(inVals_, bias, out);
   in = inVals_[0];
 
   std::shared_ptr<sum::primitive_desc> fwdPD;
-  resetFwdPD(fwdPD, inVals_, out);
+  std::shared_ptr<sum::primitive_desc> biasPD;
+  resetFwdPD(fwdPD, biasPD, inVals_, bias, out);
 
-  resetFwdPipeline(pipeline, fwdPD, inVals_, out);
+  resetFwdPipeline(pipeline, fwdPD, biasPD, inVals_, bias, out);
 }
 
 void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
@@ -79,7 +77,7 @@ void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
                                 MKLDNNMatrixPtr& wgt,
                                 MKLDNNMatrixPtr& bias,
                                 MKLDNNMatrixPtr& out) {
-  resetBwdBuffers(inGrads_, out);
+  resetBwdBuffers(inGrads_, bias, out);
   in = inGrads_[0];
 
   // backward only need share output grad to input grad
@@ -89,6 +87,20 @@ void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
       inputLayers_[i]->getOutputGrad()->setData(inGrads_[i]->getData());
     }
   }
+
+  // backward bias
+  bwdBias_ = nullptr;
+  if (bias) {
+    std::vector<float> scales(bs_, 1.0);
+    std::vector<memory::primitive_desc> srcPDs(bs_, bias->getPrimitiveDesc());
+    auto biasPD = sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs);
+    std::vector<primitive::at> srcs;
+    for (size_t i = 0; i < grads_.size(); ++i) {
+      srcs.push_back(*(grads_[i]));
+    }
+    bwdBias_.reset(new sum(biasPD, srcs, *bias));
+    pipeline.push_back(*bwdBias_);
+  }
 }
 
 void MKLDNNAddtoLayer::updateWeights(const UpdateCallback& callback) {
@@ -97,7 +109,25 @@ void MKLDNNAddtoLayer::updateWeights(const UpdateCallback& callback) {
   }
 }
 
+void MKLDNNAddtoLayer::prepareBias(MKLDNNMatrixPtr& bias,
+                                   const MatrixPtr& biasMat,
+                                   const MKLDNNMatrixPtr& out,
+                                   std::vector<MKLDNNMatrixPtr>& outs) {
+  auto pd = MKLDNNMatrix::createPrimitiveDesc(
+      {(int)layerSize_}, memory::format::x, engine_);
+  bias = MKLDNNMatrix::create(pd, biasMat);
+  outs.clear();
+  real* data = out->getData();
+  CHECK_EQ(bs_ * layerSize_, out->getElementCnt());
+  for (int i = 0; i < bs_; ++i) {
+    MatrixPtr tmp =
+        Matrix::create(data + i * layerSize_, 1, layerSize_, false, false);
+    outs.push_back(MKLDNNMatrix::create(bias->getPrimitiveDesc(), tmp));
+  }
+}
+
 void MKLDNNAddtoLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                                       MKLDNNMatrixPtr& bias,
                                        MKLDNNMatrixPtr& out) {
   inputs.resize(inputLayers_.size());
   for (size_t i = 0; i < inputs.size(); i++) {
@@ -110,12 +140,20 @@ void MKLDNNAddtoLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
   }
 
   resetOutValue(out, inputs[0]->getPrimitiveDesc());
+
+  if (biases_ && biases_->getW()) {
+    prepareBias(bias, biases_->getW(), out, vals_);
+  } else {
+    bias = nullptr;
+  }
 }
 
 void MKLDNNAddtoLayer::resetFwdPD(std::shared_ptr<sum::primitive_desc>& pd,
+                                  std::shared_ptr<sum::primitive_desc>& biasPD,
                                   std::vector<MKLDNNMatrixPtr>& inputs,
+                                  MKLDNNMatrixPtr bias,
                                   MKLDNNMatrixPtr out) {
-  std::vector<double> scales(inputs.size(), 1.0);
+  std::vector<float> scales(inputs.size(), 1.0);
   std::vector<memory::primitive_desc> srcPDs;
   for (size_t i = 0; i < inputs.size(); i++) {
     srcPDs.push_back(inputs[i]->getPrimitiveDesc());
@@ -123,12 +161,23 @@ void MKLDNNAddtoLayer::resetFwdPD(std::shared_ptr<sum::primitive_desc>& pd,
   CHECK(out);
   pd.reset(new sum::primitive_desc(out->getMemoryDesc(), scales, srcPDs));
   CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc());
+
+  biasPD = nullptr;
+  if (bias) {
+    std::vector<float> scales(2, 1.0);
+    std::vector<memory::primitive_desc> srcPDs(2, bias->getPrimitiveDesc());
+    biasPD.reset(
+        new sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs));
+    CHECK_PRIMITIVE_DESC_EQ(bias, biasPD->dst_primitive_desc());
+  }
 }
 
 void MKLDNNAddtoLayer::resetFwdPipeline(
     std::vector<primitive>& pipeline,
     std::shared_ptr<sum::primitive_desc>& pd,
+    std::shared_ptr<sum::primitive_desc>& biasPD,
     std::vector<MKLDNNMatrixPtr>& inputs,
+    MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
   std::vector<primitive::at> srcs;
   for (size_t i = 0; i < inputs.size(); i++) {
@@ -136,9 +185,23 @@ void MKLDNNAddtoLayer::resetFwdPipeline(
   }
   fwd_.reset(new sum(*pd, srcs, *out));
   pipeline.push_back(*fwd_);
+
+  fwdBias_.clear();
+  if (biasPD == nullptr || bias == nullptr) {
+    return;
+  }
+  fwdBias_.resize(vals_.size());
+  for (size_t i = 0; i < vals_.size(); ++i) {
+    std::vector<primitive::at> srcs;
+    srcs.push_back(*(vals_[i]));
+    srcs.push_back(*bias);
+    fwdBias_[i].reset(new sum(*biasPD, srcs, *vals_[i]));
+    pipeline.push_back(*fwdBias_[i]);
+  }
 }
 
 void MKLDNNAddtoLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                                       MKLDNNMatrixPtr& bias,
                                        MKLDNNMatrixPtr& out) {
   CHECK(outVal_);
   resetOutGrad(out, outVal_->getPrimitiveDesc());
@@ -149,6 +212,12 @@ void MKLDNNAddtoLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
     resetInGrad(inputs[i], inVal_->getPrimitiveDesc(), i);
     CHECK_PRIMITIVE_DESC_EQ(inputs[i], out->getPrimitiveDesc());
   }
+
+  if (biases_ && biases_->getWGrad()) {
+    prepareBias(bias, biases_->getWGrad(), out, grads_);
+  } else {
+    bias = nullptr;
+  }
 }
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNAddtoLayer.h

@@ -32,9 +32,15 @@ protected:
   // layer size == ic * ih * iw == oc * oh *ow, and can not be changed
   size_t layerSize_;
 
-  // TODO(TJ): this part has not been optimized by MKL-DNN
   std::unique_ptr<Weight> biases_;
 
+  // buffers for adding bias
+  std::vector<MKLDNNMatrixPtr> vals_;
+  std::vector<MKLDNNMatrixPtr> grads_;
+  // primitives for adding bias
+  std::vector<std::shared_ptr<mkldnn::primitive>> fwdBias_;
+  std::shared_ptr<mkldnn::primitive> bwdBias_;
+
 public:
   explicit MKLDNNAddtoLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
 
@@ -91,20 +97,34 @@ protected:
    *                    reset pipeline.
    */
   void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                       MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);
   void resetFwdPD(std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
+                  std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
                   std::vector<MKLDNNMatrixPtr>& inputs,
+                  MKLDNNMatrixPtr bias,
                   MKLDNNMatrixPtr out);
   void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
                         std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
+                        std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
                         std::vector<MKLDNNMatrixPtr>& inputs,
+                        MKLDNNMatrixPtr& bias,
                         MKLDNNMatrixPtr& out);
 
   /**
    * Backward functions: reset buffers(inputs, output, bias)
    */
   void resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                       MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);
+
+  /**
+   * prepare for bias
+   */
+  void prepareBias(MKLDNNMatrixPtr& bias,
+                   const MatrixPtr& biasMat,
+                   const MKLDNNMatrixPtr& out,
+                   std::vector<MKLDNNMatrixPtr>& outs);
 };
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNBatchNormLayer.cpp

@@ -119,7 +119,7 @@ void MKLDNNBatchNormLayer::reshape(
     int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
   reshapeInput(bs, ih, iw);
   oh = ih;
-  ow = ow;
+  ow = iw;
   // ic_ and oc can not be changed
   CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
       << "Input channel can not be changed";

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -287,7 +287,7 @@ void MKLDNNLayer::resetMergeGrad(MKLDNNMatrixPtr& out) {
     return;
   }
   CHECK(out) << "should have reset internal ouput grad";
-  std::vector<double> scales(outputMap_.size(), 1.0);
+  std::vector<float> scales(outputMap_.size(), 1.0);
   std::vector<memory::primitive_desc> srcPDs;
   std::vector<primitive::at> srcs;
   for (auto it = outputMap_.begin(); it != outputMap_.end(); ++it) {

paddle/gserver/layers/ScaleSubRegionLayer.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "ScaleSubRegionLayer.h"
+#include "paddle/utils/Stat.h"
+namespace paddle {
+
+REGISTER_LAYER(scale_sub_region, ScaleSubRegionLayer);
+
+bool ScaleSubRegionLayer::init(const LayerMap& layerMap,
+                               const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(static_cast<int>(inputLayers_.size()), 2);
+  auto& conf = config_.inputs(0).scale_sub_region_conf();
+  value_ = conf.value();
+
+  createFunction(forward_, "ScaleSubRegion", FuncConfig().set("value", value_));
+  createFunction(
+      backward_, "ScaleSubRegionGrad", FuncConfig().set("value", value_));
+
+  return true;
+}
+
+void ScaleSubRegionLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  auto in0 = getInput(0);
+  imgH_ = in0.getFrameHeight();
+  imgW_ = in0.getFrameWidth();
+  if (imgH_ == 0 || imgW_ == 0) {
+    auto& conf = config_.inputs(0).scale_sub_region_conf();
+    imgH_ = conf.image_conf().img_size_y();
+    imgW_ = conf.image_conf().img_size();
+  }
+  MatrixPtr imgV = in0.value;
+  size_t batchSize = imgV->getHeight();
+  size_t spatialSize = imgH_ * imgW_;
+  channelsNum_ = imgV->getWidth() / spatialSize;
+  shape_ = TensorShape({batchSize, channelsNum_, imgH_, imgW_});
+
+  resetOutput(batchSize, imgV->getWidth());
+  auto& out = getOutput();
+  out.setFrameHeight(imgH_);
+  out.setFrameWidth(imgW_);
+
+  MatrixPtr indicesV = getInputValue(1);
+  indicesShape_ = TensorShape({batchSize, 6});
+
+  REGISTER_TIMER_INFO("ScaleSubRegionForward", getName().c_str());
+  BufferArgs inArgs;
+  BufferArgs outArgs;
+  inArgs.addArg(*imgV, shape_);
+  inArgs.addArg(*indicesV, indicesShape_);
+  outArgs.addArg(*out.value, shape_, ASSIGN_TO);
+  forward_[0]->calc(inArgs, outArgs);
+}
+
+void ScaleSubRegionLayer::backward(const UpdateCallback& callback) {
+  REGISTER_TIMER_INFO("ScaleSubRegionBackward", getName().c_str());
+  BufferArgs inArgs;
+  BufferArgs outArgs;
+  inArgs.addArg(*getOutputGrad(), shape_);
+  inArgs.addArg(*getInputValue(1), indicesShape_);
+  outArgs.addArg(*getInputGrad(0), shape_, ADD_TO);
+  backward_[0]->calc(inArgs, outArgs);
+}
+
+}  // namespace paddle

paddle/operators/fill_constant_op.h → paddle/gserver/layers/ScaleSubRegionLayer.h

@@ -13,25 +13,40 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "paddle/framework/eigen.h"
-#include "paddle/framework/op_registry.h"
+
+#include "Layer.h"
 
 namespace paddle {
-namespace operators {
-
-template <typename Place, typename T>
-class FillConstantOpKernel : public framework::OpKernel<T> {
- public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* out = ctx.Output<framework::Tensor>("Out");
-    out->mutable_data<T>(ctx.GetPlace());
-    auto value = ctx.Attr<float>("value");
-
-    auto out_eigen = framework::EigenVector<T>::Flatten(*out);
-    auto place = ctx.GetEigenDevice<Place>();
-    out_eigen.device(place) = out_eigen.constant(static_cast<T>(value));
-  }
+
+/**
+ * \brief  For each instance, this layer can be used to multiply a value to a
+ *         specified sub continuous region. By providing start index and end
+ *         index for C/H/W, you can specify the location and shape of the
+ *         region.
+ *
+ *         input_0: Input value.
+ *         input_1: Indices value to specify the location an shape of the
+ *                  region.
+ */
+class ScaleSubRegionLayer : public Layer {
+public:
+  explicit ScaleSubRegionLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~ScaleSubRegionLayer() {}
+
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+
+  void backward(const UpdateCallback& callback = nullptr);
+
+protected:
+  TensorShape shape_;
+  TensorShape indicesShape_;
+  size_t imgH_;
+  size_t imgW_;
+  size_t channelsNum_;
+  real value_;
 };
 
-}  // namespace operators
 }  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -53,7 +53,7 @@ TEST(Operator, dot_mul) {
 TEST(Projection, context) {
   for (auto contextStart : {-5, -3, -1, 0, 3}) {
     for (auto contextLength : {1, 2, 5, 7}) {
-      for (auto batchSize : {1, 2, 5, 20, 50}) {
+      for (auto batchSize : {1, 2, 5, 20}) {
         for (auto trainablePadding : {false, true}) {
           LOG(INFO) << " contextStart=" << contextStart
                     << " contextLength=" << contextLength
@@ -585,14 +585,14 @@ TEST(Layer, maxoutLayer) {
 }
 void testFcLayer(string format, size_t nnz) {
   TestConfig config;
-  config.biasSize = 4096;
+  config.biasSize = 1024;
   config.layerConfig.set_type("fc");
-  config.layerConfig.set_size(4096);
+  config.layerConfig.set_size(1024);
   config.layerConfig.set_active_type("sigmoid");
   config.layerConfig.set_drop_rate(0.1);
 
   config.inputDefs.push_back(
-      {INPUT_DATA, "layer_0", 8192, nnz, ParaSparse(format)});
+      {INPUT_DATA, "layer_0", 2048, nnz, ParaSparse(format)});
   config.layerConfig.add_inputs();
 
   LOG(INFO) << config.inputDefs[0].sparse.sparse << " "
@@ -609,9 +609,9 @@ void testFcLayer(string format, size_t nnz) {
 }
 
 TEST(Layer, fcLayer) {
-  testFcLayer("", 4096 * 4096 * 2);
-  testFcLayer("csc", 4096 * 40);
-  testFcLayer("csr", 4096 * 40);
+  testFcLayer("", 1024 * 1024 * 2);
+  testFcLayer("csc", 1024 * 10);
+  testFcLayer("csr", 1024 * 10);
 }
 
 TEST(Layer, SelectiveFullyConnectedLayer) {
@@ -1995,7 +1995,7 @@ TEST(Layer, multibox_loss) {
 TEST(Layer, TransLayer) {
   TestConfig config;
   const int height = 128;
-  const int width = 1028;
+  const int width = 256;
   config.layerConfig.set_type("trans");
   config.layerConfig.set_size(width);
 
@@ -2358,6 +2358,38 @@ TEST(Layer, ScaleShiftLayer) {
   }
 }
 
+TEST(Layer, ScaleSubRegionLayer) {
+  const size_t batchSize = 64;
+  const size_t size = 4096;
+  TestConfig config;
+  config.layerConfig.set_type("scale_sub_region");
+  config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
+  MatrixPtr indicesV = Matrix::create(batchSize, 6, false, false);
+  auto* data = indicesV->getData();
+  for (size_t i = 0; i < batchSize; ++i) {
+    data[i * 2] = 2;
+    data[i * 2 + 1] = 4;
+    data[i * 2 + 2] = 16;
+    data[i * 2 + 3] = 32;
+    data[i * 2 + 4] = 16;
+    data[i * 2 + 5] = 32;
+  }
+  config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "indices", indicesV, {}});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ScaleSubRegionConfig* scaleSubRegionConf =
+      input->mutable_scale_sub_region_conf();
+  ImageConfig* imgConf = scaleSubRegionConf->mutable_image_conf();
+  imgConf->set_img_size(32);
+  imgConf->set_img_size_y(32);
+  imgConf->set_channels(4);
+  scaleSubRegionConf->set_value(2.0);
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "scale_sub_region", batchSize, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/gserver/tests/test_MKLDNN.cpp

@@ -269,6 +269,7 @@ void testBatchNormLayer(const testBatchNormDesc& pm) {
 TEST(MKLDNNLayer, BatchNormLayer) {
   testBatchNormLayer({4, 10, 6, 6});
   testBatchNormLayer({16, 32, 16, 16});
+  testBatchNormLayer({4, 16, 8, 10});
 }
 
 struct testImageDesc {
@@ -300,13 +301,8 @@ void testAddtoLayer(const testImageDesc& pm, const size_t nInputs) {
   TestConfig dnnConfig;
   getAddtoConfig(dnnConfig, pm, nInputs);
   dnnConfig.layerConfig.set_type("mkldnn_addto");
-  // TODO(TJ): test with bias
-  for (auto withBias : {false}) {
-    if (withBias) {
-      dnnConfig.biasSize = pm.ic * pm.ih * pm.iw;
-    } else {
-      dnnConfig.biasSize = 0;
-    }
+  for (auto withBias : {false, true}) {
+    dnnConfig.biasSize = withBias ? pm.ic * pm.ih * pm.iw : 0;
     RUN_MKLDNN_TEST_LAYER(dnnConfig, "addto", pm)
   }
 }

paddle/math/MathFunctions.cpp

@@ -206,7 +206,7 @@ double dotProduct<double>(const int n, const double* x, const double* y) {
 }
 #endif
 
-#if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML)
+#if defined(PADDLE_USE_MKLML)
 
 template <>
 void vExp<float>(const int n, const float* a, float* r) {
@@ -295,38 +295,6 @@ template void vAdd(const int n, const double* a, const double* b, double* r);
 
 #endif
 
-#ifdef PADDLE_USE_MKL
-template <>
-void vInvSqrt<float>(const int n, const float* a, float* r) {
-  vsInvSqrt(n, a, r);
-}
-
-template <>
-void vInvSqrt<double>(const int n, const double* a, double* r) {
-  vdInvSqrt(n, a, r);
-}
-
-template <>
-void vLog1p<float>(const int n, const float* a, float* r) {
-  vsLog1p(n, a, r);
-}
-
-template <>
-void vLog1p<double>(const int n, const double* a, double* r) {
-  vdLog1p(n, a, r);
-}
-
-template <>
-void vTanh<float>(const int n, const float* a, float* r) {
-  vsTanh(n, a, r);
-}
-
-template <>
-void vTanh<double>(const int n, const double* a, double* r) {
-  vdTanh(n, a, r);
-}
-#else
-
 DEFINE_MATRIX_BINARY_OP(vInvSqrt, b = 1.0f / std::sqrt(a));
 template <class T>
 void vInvSqrt(const int n, const T* a, T* r) {
@@ -357,6 +325,4 @@ template void vLog1p(const int n, const double* a, double* r);
 template void vTanh(const int n, const float* a, float* r);
 template void vTanh(const int n, const double* a, double* r);
 
-#endif
-
 }  // namespace paddle

paddle/math/MathFunctions.h

@@ -21,11 +21,6 @@ limitations under the License. */
 #include <mkl_vml_functions.h>
 #endif
 
-#ifdef PADDLE_USE_MKL
-#include <mkl.h>
-#include <mkl_lapacke.h>
-#endif
-
 #if defined(PADDLE_USE_ATLAS) || defined(PADDLE_USE_VECLIB)
 extern "C" {
 #include <cblas.h>

paddle/math/tests/TensorCheck.h

@@ -169,7 +169,7 @@ void TensorCheck(AssertEq compare,
       count++;
     }
   }
-  EXPECT_EQ(count, 0) << "There are " << count << " different element.";
+  EXPECT_EQ(count, 0) << "There are " << count << " different elements.";
 }
 
 template <typename AssertEq, typename Tensor1, typename Tensor2>

paddle/operators/accuracy_op.cu

@@ -65,7 +65,7 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
 
     size_t num_samples = inference->dims()[0];
     size_t infer_width = inference->dims()[1];
-    cudaMemset((void**)&accuracy_data, 0, sizeof(float));
+    PADDLE_ENFORCE(cudaMemset(accuracy_data, 0, sizeof(float)));
 
     if (num_samples == 0) {
       return;

paddle/operators/array_operator.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include "paddle/framework/lod_tensor_array.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+class ArrayOp : public framework::OperatorBase {
+ public:
+  ArrayOp(const std::string &type, const framework::VariableNameMap &inputs,
+          const framework::VariableNameMap &outputs,
+          const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+ protected:
+  size_t GetOffset(const framework::Scope &scope,
+                   const platform::DeviceContext &dev_ctx) const {
+    auto *i = scope.FindVar(Input("I"));
+    PADDLE_ENFORCE(i != nullptr, "I must be set");
+    auto &i_tensor = i->Get<framework::LoDTensor>();
+    PADDLE_ENFORCE_EQ(i_tensor.numel(), 1);
+    size_t offset;
+    if (platform::is_gpu_place(i_tensor.place())) {
+      // FIXME: Avoid copy from GPU to CPU
+      framework::Tensor t;
+      t.CopyFrom(i_tensor, platform::CPUPlace(), dev_ctx);
+      dev_ctx.Wait();
+      offset = static_cast<size_t>(*t.data<int64_t>());
+    } else {
+      offset = static_cast<size_t>(*i_tensor.data<int64_t>());
+    }
+    return offset;
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/array_to_lod_tensor_op.cc

@@ -140,6 +140,23 @@ class ArrayToLoDTensorInferShape : public framework::InferShapeBase {
                    "ArrayToLoDTensorOp must has input X.");
     PADDLE_ENFORCE(context->HasInput("RankTable"),
                    "ArrayToLoDTensorOp must has input RankTable.");
+    context->SetOutputDim("Out", context->GetInputDim("X"));
+  }
+};
+
+class ArrayToLoDTensorGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *grad_op = new framework::OpDescBind();
+    grad_op->SetType("lod_tensor_to_array");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetInput("RankTable", Input("RankTable"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
   }
 };
 
@@ -149,4 +166,5 @@ class ArrayToLoDTensorInferShape : public framework::InferShapeBase {
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(array_to_lod_tensor, ops::ArrayToLoDTensorOp,
                   ops::ArrayToLoDTensorOpProtoMaker,
-                  ops::ArrayToLoDTensorInferShape);
+                  ops::ArrayToLoDTensorInferShape,
+                  ops::ArrayToLoDTensorGradMaker);

paddle/operators/clip_by_norm_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/clip_by_norm_op.h"
+
+namespace paddle {
+namespace operators {
+
+class ClipByNormOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of ClipByNormOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of ClipByNormOp should not be null.");
+    auto max_norm = ctx->Attrs().Get<float>("max_norm");
+    PADDLE_ENFORCE_GT(max_norm, 0, "max_norm should be greater than 0.");
+    auto x_dims = ctx->GetInputDim("X");
+    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+class ClipByNormOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ClipByNormOpMaker(framework::OpProto* proto,
+                    framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(Tensor) The input of clip_by_norm op."
+             "The number of dimensions must be between [1, 9].");
+    AddOutput("Out",
+              "(Tensor) The output of clip_by_norm op with shape as input(X)");
+    AddAttr<float>("max_norm", "(float) The maximum norm value.");
+    AddComment(R"DOC(
+ClipByNorm operator limits the L2 norm of the input 'X' within 'max_norm'. 
+If the L2 norm of 'X' is less than or equal to 'max_norm', 'Out' will be 
+the same as 'X'. If the L2 norm of 'X' is greater than 'max_norm', 'X' will 
+be linearly scaled to make the L2 norm of 'Out' equal to 'max_norm', as 
+shown in the following formula：
+
+'Out' = 'max_norm' * 'X' / norm('X'),
+
+where norm('X') represents the L2 norm of 'X'.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(clip_by_norm, ops::ClipByNormOp,
+                             ops::ClipByNormOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    clip_by_norm, ops::ClipByNormKernel<paddle::platform::CPUPlace, float>);

paddle/operators/fill_constant_op.cu → paddle/operators/clip_by_norm_op.cu

@@ -12,13 +12,8 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#define EIGEN_USE_GPU
-#include "paddle/framework/op_registry.h"
-#include "paddle/operators/fill_constant_op.h"
+#include "paddle/operators/clip_by_norm_op.h"
 
 namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(
-    fill_constant, ops::FillConstantOpKernel<paddle::platform::GPUPlace, float>,
-    ops::FillConstantOpKernel<paddle::platform::GPUPlace, double>,
-    ops::FillConstantOpKernel<paddle::platform::GPUPlace, int>,
-    ops::FillConstantOpKernel<paddle::platform::GPUPlace, int64_t>);
+    clip_by_norm, ops::ClipByNormKernel<paddle::platform::GPUPlace, float>);

paddle/operators/clip_by_norm_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/transform.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class ClipByNormKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto max_norm = context.Attr<T>("max_norm");
+    auto* input = context.Input<Tensor>("X");
+    auto* output = context.Output<Tensor>("Out");
+    output->mutable_data<T>(context.GetPlace());
+
+    auto x = EigenVector<T>::Flatten(*input);
+    auto out = EigenVector<T>::Flatten(*output);
+    auto x_norm = x.square().sum().sqrt();
+    auto place = context.GetEigenDevice<Place>();
+
+    auto temp = (x_norm <= max_norm).template cast<T>().eval();
+    auto scaling = temp + (static_cast<T>(1) - temp) * max_norm / x_norm;
+    Eigen::array<int, 1> one_dim{{1}};
+    Eigen::DSizes<int, 1> m_dsize(input->numel());
+    out.device(place) = x * scaling.reshape(one_dim).broadcast(m_dsize);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/compare_op.cc

@@ -14,6 +14,7 @@
 
 #include "paddle/operators/compare_op.h"
 #include "paddle/framework/op_registry.h"
+
 namespace paddle {
 namespace operators {
 template <typename OpComment>
@@ -61,19 +62,34 @@ class CompareOpInferShape : public framework::InferShapeBase {
   }
 };
 
+class CompareOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  framework::OpKernelType GetKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    framework::OpKernelType kt = OperatorWithKernel::GetKernelType(ctx);
+    // CompareOp kernel's device type is decided by input tensor place
+    kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
+    return kt;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
-#define REGISTER_LOGICAL_OP(op_type, _equation)                               \
-  struct _##op_type##Comment {                                                \
-    static char type[];                                                       \
-    static char equation[];                                                   \
-  };                                                                          \
-  char _##op_type##Comment::type[]{#op_type};                                 \
-  char _##op_type##Comment::equation[]{_equation};                            \
-  REGISTER_OP_WITH_KERNEL(                                                    \
-      op_type, ::paddle::operators::CompareOpProtoMaker<_##op_type##Comment>, \
-      ::paddle::operators::CompareOpInferShape<_##op_type##Comment>,          \
+#define REGISTER_LOGICAL_OP(op_type, _equation)                      \
+  struct _##op_type##Comment {                                       \
+    static char type[];                                              \
+    static char equation[];                                          \
+  };                                                                 \
+  char _##op_type##Comment::type[]{#op_type};                        \
+  char _##op_type##Comment::equation[]{_equation};                   \
+  REGISTER_OPERATOR(                                                 \
+      op_type, ::paddle::operators::CompareOp,                       \
+      ::paddle::operators::CompareOpProtoMaker<_##op_type##Comment>, \
+      ::paddle::operators::CompareOpInferShape<_##op_type##Comment>, \
       ::paddle::framework::EmptyGradOpMaker);
 
 REGISTER_LOGICAL_OP(less_than, "Out = X < Y");

paddle/operators/fill_constant_batch_size_like_op.cc

@@ -75,10 +75,10 @@ class FillConstantBatchSizeLikeOpMaker
               "with the specified value");
     AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
     AddAttr<int>("input_dim_idx",
-                 "(int, default 0) the index of input's batch size dimension")
+                 "(int, default 0) The index of input's batch size dimension")
         .SetDefault(0);
     AddAttr<int>("output_dim_idx",
-                 "(int, default 0) the index of output's batch size dimension")
+                 "(int, default 0) The index of output's batch size dimension")
         .SetDefault(0);
     AddAttr<float>("value", "(float, default 0) The value to be filled")
         .SetDefault(0.0f);

paddle/operators/fill_constant_op.cc

@@ -12,33 +12,41 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/operators/fill_constant_op.h"
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {
 
-class FillConstantOp : public framework::OperatorWithKernel {
+class FillConstantInferShape : public framework::InferShapeBase {
  public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void operator()(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of FillConstantOp should not be null.");
     auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
-    std::vector<int64_t> shape_int64(shape.size(), 0);
-    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
-                   [](int a) { return static_cast<int64_t>(a); });
-    auto dims = framework::make_ddim(shape_int64);
-    ctx->SetOutputDim("Out", dims);
+    ctx->SetOutputDim("Out", framework::make_ddim(shape));
   }
+};
 
- protected:
-  framework::OpKernelType GetKernelType(
-      const framework::ExecutionContext &ctx) const override {
-    int data_type = ctx.Attr<int>("data_type");
-    VLOG(10) << " FillConstant data_type = " << data_type;
-    return framework::OpKernelType(static_cast<framework::DataType>(data_type),
-                                   ctx.device_context());
+class FillConstantOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto data_type = static_cast<framework::DataType>(Attr<int>("data_type"));
+    auto value = Attr<float>("value");
+    auto force_cpu = Attr<bool>("force_cpu");
+    auto &out =
+        *scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
+    out.Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+    if (force_cpu) {
+      auto cpu = platform::CPUPlace();
+      out.mutable_data(cpu, framework::ToTypeIndex(data_type));
+    } else {
+      out.mutable_data(dev_ctx.GetPlace(), framework::ToTypeIndex(data_type));
+    }
+    math::set_constant(dev_ctx, &out, value);
   }
 };
 
@@ -54,6 +62,11 @@ class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
     AddAttr<float>("value", "(float, default 0) The value to be filled")
         .SetDefault(0.0f);
+    AddAttr<bool>("force_cpu",
+                  "(bool, default false) Force fill output variable to cpu "
+                  "memory. Otherwise, fill output variable to the running "
+                  "device")
+        .SetDefault(false);
     AddOutput("Out",
               "(Tensor) Tensor of specified shape will be filled "
               "with the specified value");
@@ -69,10 +82,6 @@ Fill up a variable with specified constant value.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(fill_constant, ops::FillConstantOp,
-                             ops::FillConstantOpMaker);
-REGISTER_OP_CPU_KERNEL(
-    fill_constant, ops::FillConstantOpKernel<paddle::platform::CPUPlace, float>,
-    ops::FillConstantOpKernel<paddle::platform::CPUPlace, double>,
-    ops::FillConstantOpKernel<paddle::platform::CPUPlace, int>,
-    ops::FillConstantOpKernel<paddle::platform::CPUPlace, int64_t>);
+REGISTER_OPERATOR(fill_constant, ops::FillConstantOp,
+                  ops::FillConstantInferShape, ops::FillConstantOpMaker,
+                  paddle::framework::EmptyGradOpMaker);

paddle/operators/lod_tensor_to_array_op.cc

@@ -133,6 +133,22 @@ class LoDTensorToArrayInferVarType : public framework::VarTypeInference {
   }
 };
 
+class LoDTensorToArrayGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *grad_op = new framework::OpDescBind();
+    grad_op->SetType("array_to_lod_tensor");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetInput("RankTable", Input("RankTable"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -140,4 +156,5 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(lod_tensor_to_array, ops::LoDTensorToArrayOp,
                   ops::LoDTensorToArrayOpProtoMaker,
                   ops::LoDTensorToArrayInferShape,
-                  ops::LoDTensorToArrayInferVarType);
+                  ops::LoDTensorToArrayInferVarType,
+                  ops::LoDTensorToArrayGradMaker);

paddle/operators/lstm_op.cc

@@ -24,6 +24,11 @@ class LSTMOp : public framework::OperatorWithKernel {
   void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Input"),
                    "Input(Input) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Weight"),
+                   "Input(Weight) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Bias"),
+                   "Input(Bias) of LSTM should not be null.");
+
     PADDLE_ENFORCE(ctx->HasOutput("Hidden"),
                    "Output(Hidden) of LSTM should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Cell"),
@@ -59,11 +64,13 @@ class LSTMOp : public framework::OperatorWithKernel {
                       "The second dimension of Input(Weight) "
                       "should be 4 * %d.",
                       frame_size);
+
     auto b_dims = ctx->GetInputDim("Bias");
     PADDLE_ENFORCE_EQ(b_dims.size(), 2, "The rank of Input(Bias) should be 2.");
     PADDLE_ENFORCE_EQ(b_dims[0], 1,
                       "The first dimension of Input(Bias) should be 1.");
-    if (ctx->Attrs().Get<bool>("usePeepholes")) {
+
+    if (ctx->Attrs().Get<bool>("use_peepholes")) {
       PADDLE_ENFORCE_EQ(b_dims[1], 7 * frame_size,
                         "The second dimension of Input(Bias) should be "
                         "7 * %d if enable peepholes connection",
@@ -74,6 +81,7 @@ class LSTMOp : public framework::OperatorWithKernel {
                         "4 * %d if disable peepholes connection",
                         frame_size);
     }
+
     framework::DDim out_dims({in_dims[0], frame_size});
     ctx->SetOutputDim("Hidden", out_dims);
     ctx->SetOutputDim("Cell", out_dims);
@@ -118,14 +126,13 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Bias",
              "(Tensor) the learnable weights, which contains two parts: "
              "input-hidden bias weight and peephole connections weight if "
-             "setting `usePeepholes` True. "
-             "1. `usePeepholes = False` "
+             "setting `use_peepholes` True. "
+             "1. `use_peepholes = False` "
              " - The shape is (1 x 4D). "
              " - Bias = {b_c, b_i, b_f, b_o}."
-             "2. `usePeepholes = True` "
+             "2. `use_peepholes = True` "
              " - The shape is (1 x 7D). "
-             " - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.")
-        .AsDispensable();
+             " - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.");
     AddOutput("Hidden",
               "(LoDTensor) the hidden state of LSTM operator. "
               "The shape is (T x D), and lod is the same with the `Input`.");
@@ -145,29 +152,32 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
               "(LoDTensor) This LoDTensor is obtained in the forward and used "
               "in the backward.")
         .AsIntermediate();
-    AddAttr<bool>("usePeepholes",
-                  "(bool, default True) "
+    AddAttr<bool>("use_peepholes",
+                  "(bool, defalut: True) "
                   "whether to enable diagonal/peephole connections.")
         .SetDefault(true);
-    AddAttr<bool>("isReverse",
-                  "(bool, default False) "
+    AddAttr<bool>("is_reverse",
+                  "(bool, defalut: False) "
                   "whether to compute reversed LSTM.")
         .SetDefault(false);
     AddAttr<std::string>(
-        "gateActivation",
-        "(string, default sigmoid)"
+        "gate_activation",
+        "(string, default: sigmoid)"
         "The activation for input gate, forget gate and output "
         "gate, `sigmoid` by default.")
-        .SetDefault("sigmoid");
-    AddAttr<std::string>("cellActivation",
-                         "(string, default tanh)"
+        .SetDefault("sigmoid")
+        .InEnum({"sigmoid", "tanh", "relu", "identity"});
+    AddAttr<std::string>("cell_activation",
+                         "(string, default: tanh)"
                          "The activation for cell output, `tanh` by defalut.")
-        .SetDefault("tanh");
-    AddAttr<std::string>("candidateActivation",
-                         "(string, default tanh)"
+        .SetDefault("tanh")
+        .InEnum({"sigmoid", "tanh", "relu", "identity"});
+    AddAttr<std::string>("candidate_activation",
+                         "(string, default: tanh)"
                          "The activation for candidate hidden state, "
                          "`tanh` by default.")
-        .SetDefault("tanh");
+        .SetDefault("tanh")
+        .InEnum({"sigmoid", "tanh", "relu", "identity"});
     AddComment(R"DOC(
 Long-Short Term Memory (LSTM) Operator.
 
@@ -203,7 +213,7 @@ are the cell input and cell output activation functions and `tanh` is usually
 used for them. \f$\tilde{c_t}\f$ is also called candidate hidden state,
 which is computed based on the current input and the previous hidden state.
 
-Set usePeepholes False to disable peephole connection 
+Set `use_peepholes` False to disable peephole connection 
 (http://www.bioinf.jku.at/publications/older/2604.pdf). The formula
 is omitted here.
 
@@ -226,23 +236,27 @@ class LSTMGradOp : public framework::OperatorWithKernel {
                    "Input(Hidden) of LSTM should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Cell"),
                    "Input(Cell) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Weight"),
+                   "Input(Weight) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Bias"),
+                   "Input(Bias) of LSTM should not be null.");
 
     PADDLE_ENFORCE(ctx->HasInput("BatchGate"),
                    "Input(BatchGate) of LSTM should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("BatchCellPreAct"),
                    "Input(BatchGate) of LSTM should not be null.");
 
-    auto in_g_name = framework::GradVarName("Input");
-    if (ctx->HasOutput(in_g_name))
-      ctx->SetOutputDim(in_g_name, ctx->GetInputDim("Input"));
-
-    auto w_g_name = framework::GradVarName("Weight");
-    if (ctx->HasOutput(w_g_name))
-      ctx->SetOutputDim(w_g_name, ctx->GetInputDim("Weight"));
-
-    auto b_g_name = framework::GradVarName("Bias");
-    if (ctx->HasOutput(b_g_name))
-      ctx->SetOutputDim(b_g_name, ctx->GetInputDim("Bias"));
+    auto SetOutGradDim = [&ctx](const std::string& name) {
+      auto g_name = framework::GradVarName(name);
+      if (ctx->HasOutput(g_name))
+        ctx->SetOutputDim(g_name, ctx->GetInputDim(name));
+    };
+
+    SetOutGradDim("Input");
+    SetOutGradDim("Weight");
+    SetOutGradDim("Bias");
+    SetOutGradDim("H0");
+    SetOutGradDim("C0");
   }
 
  protected:

paddle/operators/lstm_op.h

@@ -28,6 +28,15 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
+template <typename Place, typename T>
+inline void ReorderInitState(const platform::DeviceContext& ctx,
+                             const framework::Tensor& src, const size_t* index,
+                             framework::Tensor* dst, bool indexed_src) {
+  math::CopyMatrixRowsFunctor<Place, T> row_shuffle;
+  dst->mutable_data<T>(src.dims(), ctx.GetPlace());
+  row_shuffle(ctx, src, index, *dst, indexed_src);
+}
+
 template <typename Place, typename T>
 class LSTMKernel : public framework::OpKernel<T> {
  public:
@@ -36,6 +45,9 @@ class LSTMKernel : public framework::OpKernel<T> {
     auto* weight = ctx.Input<Tensor>("Weight");
     auto* bias = ctx.Input<Tensor>("Bias");
 
+    auto* hidden_t0 = ctx.Input<Tensor>("H0");
+    auto* cell_t0 = ctx.Input<Tensor>("C0");
+
     auto* batch_gate = ctx.Output<LoDTensor>("BatchGate");
     batch_gate->mutable_data<T>(ctx.GetPlace());
     auto* hidden_out = ctx.Output<LoDTensor>("Hidden");
@@ -43,12 +55,7 @@ class LSTMKernel : public framework::OpKernel<T> {
     auto* cell_out = ctx.Output<LoDTensor>("Cell");
     cell_out->mutable_data<T>(ctx.GetPlace());
 
-    // Now the function ShareLoD in InferShape is not implemented.
-    // So copy LoD here.
-    ctx.ShareLoD("Input", "Hidden");
-    ctx.ShareLoD("Input", "Cell");
-
-    bool is_reverse = ctx.Attr<bool>("isReverse");
+    bool is_reverse = ctx.Attr<bool>("is_reverse");
     math::LoDTensor2BatchFunctor<Place, T> to_batch;
     auto& device_ctx = ctx.device_context();
     to_batch(device_ctx, *input, *batch_gate, true, is_reverse);
@@ -71,7 +78,7 @@ class LSTMKernel : public framework::OpKernel<T> {
     }
 
     math::LstmMetaValue<T> lstm_value;
-    if (bias) {
+    if (bias && ctx.Attr<bool>("use_peepholes")) {
       T* bias_data = const_cast<T*>(bias->data<T>());
       // the code style in LstmMetaValue will be updated later.
 
@@ -84,6 +91,16 @@ class LSTMKernel : public framework::OpKernel<T> {
       lstm_value.checkOg = nullptr;
     }
     lstm_value.prevStateValue = nullptr;
+    Tensor ordered_c0;
+    const size_t* order = batch_gate->lod()[2].data();
+    if (cell_t0) {
+      // Since the batch computing for LSTM reorders the input sequence
+      // according to their length. The initialized cell state also needs
+      // to reorder.
+      ReorderInitState<Place, T>(device_ctx, *cell_t0, order, &ordered_c0,
+                                 true);
+      lstm_value.prevStateValue = ordered_c0.data<T>();
+    }
 
     // Use the local variable as here.
     LoDTensor batch_hidden, batch_cell;
@@ -94,9 +111,9 @@ class LSTMKernel : public framework::OpKernel<T> {
 
     auto batch_starts = batch_gate->lod()[0];
     size_t num_batch = batch_starts.size() - 1;
-    auto gate_act = ctx.Attr<std::string>("gateActivation");
-    auto cell_act = ctx.Attr<std::string>("cellActivation");
-    auto cand_act = ctx.Attr<std::string>("candidateActivation");
+    auto gate_act = ctx.Attr<std::string>("gate_activation");
+    auto cell_act = ctx.Attr<std::string>("cell_activation");
+    auto cand_act = ctx.Attr<std::string>("candidate_activation");
 
     for (size_t n = 0; n < num_batch; n++) {
       int bstart = static_cast<int>(batch_starts[n]);
@@ -109,15 +126,28 @@ class LSTMKernel : public framework::OpKernel<T> {
 
       int cur_batch_size = bend - bstart;
 
-      if (n != 0) {
+      if (n > 0) {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_hidden_t = batch_hidden.Slice(pre_h_start, pre_h_end);
         math::matmul<Place, T>(device_ctx, pre_hidden_t, false, *weight, false,
                                static_cast<T>(1.0), &gate_t,
                                static_cast<T>(1.0));
+      } else if (hidden_t0) {
+        // If n == 0 and there is no initialized hidden state, that is to say
+        // the H0 is zeros, the calculation W_h * H0 will be skiped.
+        // If n == 0 and there is initialized hidden state, calculate W_h * H0.
+
+        // Since the batch computing for LSTM reorders the input sequence
+        // according to their length. The initialized hidden state also needs
+        // to reorder.
+        Tensor ordered_h0;
+        ReorderInitState<Place, T>(device_ctx, *hidden_t0, order, &ordered_h0,
+                                   true);
+        math::matmul<Place, T>(device_ctx, ordered_h0, false, *weight, false,
+                               static_cast<T>(1.0), &gate_t,
+                               static_cast<T>(1.0));
       }
-      // else if : FIXME support the initial hidden and cell
 
       lstm_value.gateValue = gate_t.data<T>();
       lstm_value.outputValue = out_t.data<T>();
@@ -160,6 +190,12 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     auto* weight_g = ctx.Output<Tensor>(framework::GradVarName("Weight"));
     auto* bias_g = ctx.Output<Tensor>(framework::GradVarName("Bias"));
 
+    auto* h0 = ctx.Input<Tensor>("H0");
+    auto* c0 = ctx.Input<Tensor>("C0");
+
+    auto* h0_g = ctx.Output<Tensor>(framework::GradVarName("H0"));
+    auto* c0_g = ctx.Output<Tensor>(framework::GradVarName("C0"));
+
     auto& device_ctx = ctx.device_context();
     math::SetConstant<Place, T> zero;
     if (weight_g) {
@@ -167,13 +203,25 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       zero(device_ctx, weight_g, static_cast<T>(0.0));
     }
 
+    // ordered_h0/c0 is the reordered hidden/cell initialization.
+    // ordered_h0_g/c0_g is the reordered gradient of hidden/cell
+    // initialization.
+    Tensor ordered_h0, ordered_c0, ordered_h0_g, ordered_c0_g;
+    const size_t* order = batch_gate->lod()[2].data();
+    if (c0) {
+      ReorderInitState<Place, T>(device_ctx, *c0, order, &ordered_c0, true);
+    }
+    if (c0 && c0_g) {
+      ordered_c0_g.mutable_data<T>(c0_g->dims(), ctx.GetPlace());
+    }
+
     auto in_dims = input->dims();
     auto out_dims = hidden_g->dims();
     int frame_size = static_cast<int>(in_dims[1] / 4);
     PADDLE_ENFORCE_EQ(frame_size, out_dims[1]);
 
     math::LstmMetaValue<T> lstm_value;
-    if (bias) {
+    if (bias && ctx.Attr<bool>("use_peepholes")) {
       T* bias_data = const_cast<T*>(bias->data<T>());
       lstm_value.checkIg = bias_data + 4 * frame_size;
       lstm_value.checkFg = lstm_value.checkIg + frame_size;
@@ -185,9 +233,13 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     }
 
     math::LstmMetaGrad<T> lstm_grad;
+
     if (bias && bias_g) {
-      T* bias_g_data = const_cast<T*>(bias_g->mutable_data<T>(ctx.GetPlace()));
+      bias_g->mutable_data<T>(ctx.GetPlace());
       zero(device_ctx, bias_g, static_cast<T>(0.0));
+    }
+    if (bias && bias_g && ctx.Attr<bool>("use_peepholes")) {
+      T* bias_g_data = bias_g->data<T>();
       lstm_grad.checkIgGrad = bias_g_data + 4 * frame_size;
       lstm_grad.checkFgGrad = lstm_grad.checkIgGrad + frame_size;
       lstm_grad.checkOgGrad = lstm_grad.checkFgGrad + frame_size;
@@ -199,36 +251,30 @@ class LSTMGradKernel : public framework::OpKernel<T> {
 
     math::LoDTensor2BatchFunctor<Place, T> to_batch;
 
-    // use the local variable as here.
-    LoDTensor batch_hidden;
-    batch_hidden.mutable_data<T>(out_dims, ctx.GetPlace());
-    batch_hidden.set_lod(batch_gate->lod());
-    to_batch(device_ctx, *hidden_out, batch_hidden, false);
+    auto ToBatch = [&batch_gate, &to_batch](
+        const platform::DeviceContext& ctx, const framework::LoDTensor& src,
+        const framework::DDim& dims, framework::LoDTensor& dst) {
+      dst.mutable_data<T>(dims, ctx.GetPlace());
+      dst.set_lod(batch_gate->lod());
+      to_batch(ctx, src, dst, false);
+    };
 
-    LoDTensor batch_hidden_g;
-    batch_hidden_g.mutable_data<T>(out_dims, ctx.GetPlace());
-    batch_hidden_g.set_lod(batch_gate->lod());
-    to_batch(device_ctx, *hidden_g, batch_hidden_g, false);
+    LoDTensor batch_hidden, batch_hidden_g, batch_cell;
+    ToBatch(device_ctx, *hidden_out, out_dims, batch_hidden);
+    ToBatch(device_ctx, *hidden_g, out_dims, batch_hidden_g);
+    ToBatch(device_ctx, *cell_out, out_dims, batch_cell);
 
-    LoDTensor batch_cell;
-    batch_cell.mutable_data<T>(out_dims, ctx.GetPlace());
-    batch_cell.set_lod(batch_gate->lod());
-    to_batch(device_ctx, *cell_out, batch_cell, false);
-
-    LoDTensor batch_cell_g;
+    LoDTensor batch_cell_g, batch_gate_g;
     batch_cell_g.mutable_data<T>(out_dims, ctx.GetPlace());
-    batch_cell_g.set_lod(batch_gate->lod());
     // TODO(qingqing) support the case output cell has gradient.
     // to_batch(device_ctx, *cell_g, batch_cell_g, false);
     zero(device_ctx, &batch_cell_g, static_cast<T>(0.0));
-
-    LoDTensor batch_gate_g;
     batch_gate_g.mutable_data<T>(batch_gate->dims(), ctx.GetPlace());
     batch_gate_g.set_lod(batch_gate->lod());
 
-    auto gate_act = ctx.Attr<std::string>("gateActivation");
-    auto cell_act = ctx.Attr<std::string>("cellActivation");
-    auto cand_act = ctx.Attr<std::string>("candidateActivation");
+    auto gate_act = ctx.Attr<std::string>("gate_activation");
+    auto cell_act = ctx.Attr<std::string>("cell_activation");
+    auto cand_act = ctx.Attr<std::string>("candidate_activation");
 
     auto batch_starts = batch_gate->lod()[0];
     size_t num_batch = batch_starts.size() - 1;
@@ -250,15 +296,15 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       lstm_grad.gateGrad = gate_g.data<T>();
       lstm_grad.outputGrad = out_g.data<T>();
 
-      if (n) {
+      if (n > 0) {
         int bstart_pre = static_cast<int>(batch_starts[n - 1]);
         Tensor cell_pre = batch_cell.Slice(bstart_pre, bstart);
         Tensor cell_pre_g = batch_cell_g.Slice(bstart_pre, bstart);
         lstm_value.prevStateValue = cell_pre.data<T>();
         lstm_grad.prevStateGrad = cell_pre_g.data<T>();
       } else {
-        lstm_value.prevStateValue = nullptr;
-        lstm_grad.prevStateGrad = nullptr;
+        lstm_value.prevStateValue = c0 ? ordered_c0.data<T>() : nullptr;
+        lstm_grad.prevStateGrad = c0_g ? ordered_c0_g.data<T>() : nullptr;
       }
 
       int cur_batch_size = bend - bstart;
@@ -266,7 +312,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
           device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size,
           gate_act, cell_act, cand_act);
 
-      if (n != 0) {
+      if (n > 0) {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_hidden_g = batch_hidden_g.Slice(pre_h_start, pre_h_end);
@@ -280,6 +326,19 @@ class LSTMGradKernel : public framework::OpKernel<T> {
                                  static_cast<T>(1.0), weight_g,
                                  static_cast<T>(1.0));
         }
+      } else {
+        if (h0 && weight_g) {
+          ReorderInitState<Place, T>(device_ctx, *h0, order, &ordered_h0, true);
+          math::matmul<Place, T>(device_ctx, ordered_h0, true, gate_g, false,
+                                 static_cast<T>(1.0), weight_g,
+                                 static_cast<T>(1.0));
+        }
+        if (h0 && h0_g) {
+          ordered_h0_g.mutable_data<T>(h0_g->dims(), ctx.GetPlace());
+          math::matmul<Place, T>(device_ctx, gate_g, false, *weight, true,
+                                 static_cast<T>(1.0), &ordered_h0_g,
+                                 static_cast<T>(0.0));
+        }
       }
     }
 
@@ -302,6 +361,13 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       math::gemv<Place, T>(device_ctx, true, m, n, 1., batch_gate_g.data<T>(),
                            ones.data<T>(), 0., bias_g->data<T>());
     }
+
+    if (h0 && h0_g) {
+      ReorderInitState<Place, T>(device_ctx, ordered_h0_g, order, h0_g, false);
+    }
+    if (c0 && c0_g) {
+      ReorderInitState<Place, T>(device_ctx, ordered_c0_g, order, c0_g, false);
+    }
   }
 };
 

paddle/operators/lstm_unit_op.cc

@@ -34,10 +34,10 @@ class LstmUnitOp : public framework::OperatorWithKernel {
     auto c_prev_dims = ctx->GetInputDim("C_prev");
 
     PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2.");
-    PADDLE_ENFORCE(x_dims[0] == c_prev_dims[0],
-                   "Batch size of inputs and states must be equal");
-    PADDLE_ENFORCE(x_dims[1] == c_prev_dims[1] * 4,
-                   "Dimension of FC should equal to prev state * 4");
+    PADDLE_ENFORCE_EQ(x_dims[0], c_prev_dims[0],
+                      "Batch size of inputs and states must be equal");
+    PADDLE_ENFORCE_EQ(x_dims[1], c_prev_dims[1] * 4,
+                      "Dimension of FC should equal to prev state * 4");
 
     int b_size = c_prev_dims[0];  // batch size
     int s_dim = c_prev_dims[1];   // state dim

paddle/operators/math/detail/lstm_cpu_kernel.h

@@ -52,9 +52,9 @@ void naive_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value,
     rValueIg = valueIg[i];
     rValueFg = valueFg[i];
     rValueOg = valueOg[i];
-    rCheckI = value.checkIg[i];
-    rCheckF = value.checkFg[i];
-    rCheckO = value.checkOg[i];
+    rCheckI = value.checkIg ? value.checkIg[i] : 0;
+    rCheckF = value.checkFg ? value.checkFg[i] : 0;
+    rCheckO = value.checkOg ? value.checkOg[i] : 0;
 
     if (value.prevStateValue) {
       rPrevState = value.prevStateValue[i];
@@ -114,9 +114,9 @@ void naive_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
     rValueIg = valueIg[i];
     rValueFg = valueFg[i];
     rValueOg = valueOg[i];
-    rCheckI = value.checkIg[i];
-    rCheckF = value.checkFg[i];
-    rCheckO = value.checkOg[i];
+    rCheckI = value.checkIg ? value.checkIg[i] : 0;
+    rCheckF = value.checkFg ? value.checkFg[i] : 0;
+    rCheckO = value.checkOg ? value.checkOg[i] : 0;
     rState = value.stateValue[i];
     rStateAtv = value.stateActiveValue[i];
     rOutputGrad = grad.outputGrad[i];
@@ -155,9 +155,9 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value, int frameSize,
   __m256 rValueIg;
   __m256 rValueFg;
   __m256 rValueOg;
-  __m256 rCheckI;
-  __m256 rCheckF;
-  __m256 rCheckO;
+  __m256 rCheckI = _mm256_set1_ps(0.0f);
+  __m256 rCheckF = _mm256_set1_ps(0.0f);
+  __m256 rCheckO = _mm256_set1_ps(0.0f);
   __m256 rState;
   __m256 rPrevState = _mm256_set1_ps(0.0f);
   __m256 rStateAtv;
@@ -173,9 +173,11 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value, int frameSize,
     rValueIg = valueIg[i];
     rValueFg = valueFg[i];
     rValueOg = valueOg[i];
-    rCheckI = ((__m256 *)value.checkIg)[i];
-    rCheckF = ((__m256 *)value.checkFg)[i];
-    rCheckO = ((__m256 *)value.checkOg)[i];
+    if (value.checkIg) {
+      rCheckI = ((__m256 *)value.checkIg)[i];
+      rCheckF = ((__m256 *)value.checkFg)[i];
+      rCheckO = ((__m256 *)value.checkOg)[i];
+    }
 
     if (value.prevStateValue) {
       rPrevState = ((__m256 *)value.prevStateValue)[i];
@@ -216,9 +218,9 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
   __m256 rState;
   __m256 rStateAtv;
   __m256 rOutputGrad;
-  __m256 rCheckI;
-  __m256 rCheckF;
-  __m256 rCheckO;
+  __m256 rCheckI = _mm256_set1_ps(0.0f);
+  __m256 rCheckF = _mm256_set1_ps(0.0f);
+  __m256 rCheckO = _mm256_set1_ps(0.0f);
   __m256 rCheckIGrad;
   __m256 rCheckFGrad;
   __m256 rCheckOGrad;
@@ -237,9 +239,11 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
     rValueIg = valueIg[i];
     rValueFg = valueFg[i];
     rValueOg = valueOg[i];
-    rCheckI = ((__m256 *)value.checkIg)[i];
-    rCheckF = ((__m256 *)value.checkFg)[i];
-    rCheckO = ((__m256 *)value.checkOg)[i];
+    if (value.checkIg) {
+      rCheckI = ((__m256 *)value.checkIg)[i];
+      rCheckF = ((__m256 *)value.checkFg)[i];
+      rCheckO = ((__m256 *)value.checkOg)[i];
+    }
     rState = ((__m256 *)value.stateValue)[i];
     rStateAtv = ((__m256 *)value.stateActiveValue)[i];
     rOutputGrad = ((__m256 *)grad.outputGrad)[i];

paddle/operators/math/detail/lstm_gpu_kernel.h

@@ -55,9 +55,10 @@ __global__ void KeLstmForward(Op op, LstmMetaValue<T> value, int frameSize,
   T rValueIg;
   T rValueFg;
   T rValueOg;
-  T rCheckI = value.checkIg[frameIdx];
-  T rCheckF = value.checkFg[frameIdx];
-  T rCheckO = value.checkOg[frameIdx];
+
+  T rCheckI = value.checkIg ? value.checkIg[frameIdx] : 0;
+  T rCheckF = value.checkFg ? value.checkFg[frameIdx] : 0;
+  T rCheckO = value.checkOg ? value.checkOg[frameIdx] : 0;
 
   rValueIn = value.gateValue[frameIdx];
   rValueIg = value.gateValue[frameIdx + frameSize];
@@ -121,9 +122,10 @@ __global__ void KeLstmBackward(Op op, LstmMetaValue<T> value,
   T rStateGrad;
   T rStateAtv;
   T rOutputGrad;
-  T rCheckI = value.checkIg[frameIdx];
-  T rCheckF = value.checkFg[frameIdx];
-  T rCheckO = value.checkOg[frameIdx];
+  T rCheckI = value.checkIg ? value.checkIg[frameIdx] : 0;
+  T rCheckF = value.checkFg ? value.checkFg[frameIdx] : 0;
+  T rCheckO = value.checkOg ? value.checkOg[frameIdx] : 0;
+
   T rCheckIGrad;
   T rCheckFGrad;
   T rCheckOGrad;

paddle/operators/math/math_function.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/math/math_function.h"
+#include "paddle/framework/data_type.h"
 
 namespace paddle {
 namespace operators {
@@ -233,6 +234,52 @@ void gemv<platform::CPUPlace, double>(const platform::DeviceContext& context,
 
 template struct SetConstant<platform::CPUPlace, float>;
 
+struct TensorSetConstant {
+  TensorSetConstant(framework::Tensor* tensor, float value)
+      : tensor_(tensor), value_(value) {}
+  template <typename T>
+  void operator()() const {
+    auto cpu = platform::CPUPlace();
+    auto* begin = tensor_->mutable_data<T>(cpu);
+    std::fill(begin, begin + tensor_->numel(), static_cast<T>(value_));
+  }
+  framework::Tensor* tensor_;
+  float value_;
+};
+
+template <>
+void set_constant_with_place<platform::CPUPlace>(
+    const platform::DeviceContext& context, framework::Tensor* tensor,
+    float value) {
+  framework::VisitDataType(framework::ToDataType(tensor->type()),
+                           TensorSetConstant(tensor, value));
+}
+
+struct TensorSetConstantWithPlace : public boost::static_visitor<void> {
+  TensorSetConstantWithPlace(const platform::DeviceContext& context,
+                             framework::Tensor* tensor, float value)
+      : context_(context), tensor_(tensor), value_(value) {}
+
+  template <typename Place>
+  void operator()(Place place) const {
+    set_constant_with_place<Place>(context_, tensor_, value_);
+  }
+
+  const platform::DeviceContext& context_;
+  framework::Tensor* tensor_;
+  float value_;
+};
+
+void set_constant(const platform::DeviceContext& context,
+                  framework::Tensor* tensor, float value) {
+  TensorSetConstantWithPlace func(context, tensor, value);
+#ifdef PADDLE_WITH_CUDA
+  tensor->place().apply_visitor(func);
+#else
+  func(platform::CPUPlace());
+#endif
+}
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/math_function.cu

@@ -12,6 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include "paddle/framework/data_type.h"
 #include "paddle/operators/math/math_function.h"
 
 namespace paddle {
@@ -232,6 +233,30 @@ void gemv<platform::GPUPlace, double>(const platform::DeviceContext& context,
 
 template struct SetConstant<platform::GPUPlace, float>;
 
+struct TensorSetConstant {
+  TensorSetConstant(const platform::DeviceContext& context,
+                    framework::Tensor* tensor, float value)
+      : context_(context), tensor_(tensor), value_(value) {}
+
+  template <typename T>
+  void operator()() const {
+    SetConstant<platform::GPUPlace, T> functor;
+    functor(context_, tensor_, static_cast<T>(value_));
+  }
+
+  const platform::DeviceContext& context_;
+  framework::Tensor* tensor_;
+  float value_;
+};
+
+template <>
+void set_constant_with_place<platform::GPUPlace>(
+    const platform::DeviceContext& context, framework::Tensor* tensor,
+    float value) {
+  framework::VisitDataType(framework::ToDataType(tensor->type()),
+                           TensorSetConstant(context, tensor, value));
+}
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/math_function.h

@@ -19,11 +19,6 @@ limitations under the License. */
 #include <mkl_vml_functions.h>
 #endif
 
-#ifdef PADDLE_USE_MKL
-#include <mkl.h>
-#include <mkl_lapacke.h>
-#endif
-
 #ifdef PADDLE_USE_ATLAS
 extern "C" {
 #include <cblas.h>
@@ -108,6 +103,13 @@ struct SetConstant {
   }
 };
 
+template <typename Place>
+void set_constant_with_place(const platform::DeviceContext& context,
+                             framework::Tensor* tensor, float value);
+
+void set_constant(const platform::DeviceContext& context,
+                  framework::Tensor* tensor, float value);
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/math_function_test.cc

@@ -139,3 +139,15 @@ TEST(math_function, gemv) {
   GemvTest<float>(12, 7, true);
   GemvTest<double>(7, 9, true);
 }
+
+TEST(math_funciton, set_constant) {
+  paddle::framework::Tensor t;
+  t.Resize({10, 10});
+  t.mutable_data<int>(paddle::platform::CPUPlace());
+  auto* ctx = new paddle::platform::CPUDeviceContext();
+  paddle::operators::math::set_constant(*ctx, &t, 10);
+  for (int64_t i = 0; i < t.numel(); ++i) {
+    PADDLE_ENFORCE_EQ(10, t.data<int>()[i]);
+  }
+  delete ctx;
+}

paddle/operators/math/sequence2batch.cc

@@ -22,8 +22,8 @@ template <typename T>
 class CopyMatrixRowsFunctor<platform::CPUPlace, T> {
  public:
   void operator()(const platform::DeviceContext& context,
-                  const framework::LoDTensor& src, const size_t* index,
-                  framework::LoDTensor& dst, bool is_src_index) {
+                  const framework::Tensor& src, const size_t* index,
+                  framework::Tensor& dst, bool is_src_index) {
     auto src_dims = src.dims();
     auto dst_dims = dst.dims();
     PADDLE_ENFORCE_EQ(src_dims.size(), 2UL,

paddle/operators/math/sequence2batch.cu

@@ -41,8 +41,8 @@ template <typename T>
 class CopyMatrixRowsFunctor<platform::GPUPlace, T> {
  public:
   void operator()(const platform::DeviceContext& context,
-                  const framework::LoDTensor& src, const size_t* index,
-                  framework::LoDTensor& dst, bool is_src_index) {
+                  const framework::Tensor& src, const size_t* index,
+                  framework::Tensor& dst, bool is_src_index) {
     auto src_dims = src.dims();
     auto dst_dims = dst.dims();
     PADDLE_ENFORCE_EQ(src_dims.size(), 2,

paddle/operators/math/sequence2batch.h

@@ -30,8 +30,8 @@ class CopyMatrixRowsFunctor {
   // copy the input src to the indexed rows of output dst.
   // The indexed rows are based on the input index.
   void operator()(const platform::DeviceContext& context,
-                  const framework::LoDTensor& src, const size_t* index,
-                  framework::LoDTensor& dst, bool is_src_index);
+                  const framework::Tensor& src, const size_t* index,
+                  framework::Tensor& dst, bool is_src_index);
 };
 
 template <typename Place, typename T>
@@ -57,7 +57,7 @@ class LoDTensor2BatchFunctor {
                   bool is_reverse = false) const {
     if (!is_cal_batch_lod) {
       auto lods = batch.lod();
-      PADDLE_ENFORCE_EQ(lods.size(), 2UL);
+      PADDLE_ENFORCE_GT(lods.size(), 2UL);
       PADDLE_ENFORCE_EQ(lods[1].size(),
                         static_cast<size_t>(lod_tensor.dims()[0]));
       CopyMatrixRowsFunctor<Place, T> to_batch;
@@ -66,8 +66,8 @@ class LoDTensor2BatchFunctor {
     }
 
     auto lods = lod_tensor.lod();
-    PADDLE_ENFORCE_EQ(lods.size(), 1UL, "Only support one level sequence now.");
     auto lod = lods[0];
+    PADDLE_ENFORCE_EQ(lods.size(), 1UL, "Only support one level sequence now.");
 
     std::vector<SeqInfo> seq_info;
     for (size_t seq_id = 0; seq_id < lod.size() - 1; ++seq_id) {
@@ -78,8 +78,7 @@ class LoDTensor2BatchFunctor {
     std::sort(seq_info.begin(), seq_info.end(),
               [](SeqInfo a, SeqInfo b) { return a.length > b.length; });
 
-    // calculate the start position of each batch
-    // (numBatch equal the maxLength of sequences)
+    // Calculate the start position of each batch.
     // example:  sequences = {s0, s1, s2}
     //           s0: 0 0 0 0, s1: 1 1 1 1 1, s2: 2 2 2
     //           num_batch = 5,
@@ -95,19 +94,25 @@ class LoDTensor2BatchFunctor {
     //                                6, 2, 11,
     //                                7, 3,
     //                                8}
-    // The batch number represents batch size after rearranging the
+    //           seq_order = {1, 0, 2}, the sort order.
+    //               where 1 is the second sequence,
+    //                     0 is the first sequence,
+    //                     2 is the third sequence.
+    // The num_batch represents batch size after rearranging the
     // input LodTensor. It is also the maximum length of input sequence.
 
     paddle::framework::LoD batch_lods;
     batch_lods.emplace_back(std::vector<size_t>{0});
     batch_lods.emplace_back(std::vector<size_t>{0});
+    batch_lods.emplace_back(std::vector<size_t>{0});
 
     // batch_lods[0] is the start positions for batch LoDTensor
     int num_batch = seq_info[0].length;
     batch_lods[0].resize(static_cast<size_t>(num_batch + 1));
     // batch_lods[1] is the raw index in the input LoDTensor
-    auto dims = lod_tensor.dims();
-    batch_lods[1].resize(static_cast<size_t>(dims[0]));
+    batch_lods[1].resize(static_cast<size_t>(lod_tensor.dims()[0]));
+    // batch_lods[2] is the sort order for the input LoDTensor.
+    batch_lods[2].resize(seq_info.size());
 
     size_t* batch_starts = batch_lods[0].data();
     size_t* seq2batch_idx = batch_lods[1].data();
@@ -127,6 +132,10 @@ class LoDTensor2BatchFunctor {
       }
       batch_starts[n + 1] = static_cast<size_t>(batch_id);
     }
+    size_t* seq_order = batch_lods[2].data();
+    for (size_t i = 0; i < seq_info.size(); ++i) {
+      seq_order[i] = seq_info[i].seq_idx;
+    }
     batch.set_lod(batch_lods);
 
     CopyMatrixRowsFunctor<Place, T> to_batch;
@@ -141,8 +150,7 @@ class Batch2LoDTensorFunctor {
                   const framework::LoDTensor& batch,
                   framework::LoDTensor& lod_tensor) const {
     auto in_lod = batch.lod();
-    PADDLE_ENFORCE_EQ(in_lod.size(), 2UL,
-                      "The LoD size of input `batch` should be 2.");
+    PADDLE_ENFORCE_GT(in_lod.size(), 2UL);
     PADDLE_ENFORCE_EQ(in_lod[1].size(),
                       static_cast<size_t>(lod_tensor.dims()[0]));
     CopyMatrixRowsFunctor<Place, T> to_seq;

paddle/operators/mean_op.cc

@@ -51,6 +51,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
 
   void InferShape(framework::InferShapeContext* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", framework::GradVarName("X"));
   }
 };
 

paddle/operators/pool_cudnn_op.cu

@@ -37,11 +37,11 @@ class PoolCudnnOpKernel : public framework::OpKernel<T> {
     const T *input_data = input->data<T>();
     T *output_data = output->mutable_data<T>(ctx.GetPlace());
 
-    std::string pooling_type = ctx.Attr<std::string>("poolingType");
+    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
     std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
-    if (ctx.Attr<bool>("globalPooling")) {
+    if (ctx.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(input->dims()[i + 2]);
@@ -92,12 +92,12 @@ class PoolCudnnGradOpKernel : public framework::OpKernel<T> {
         ctx.Input<Tensor>(framework::GradVarName("Out"));
     Tensor *input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
 
-    std::string pooling_type = ctx.Attr<std::string>("poolingType");
+    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
     std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
 
-    if (ctx.Attr<bool>("globalPooling")) {
+    if (ctx.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(input->dims()[i + 2]);

paddle/operators/pool_op.cc

@@ -29,7 +29,7 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
 
   auto in_x_dims = ctx->GetInputDim("X");
 
-  std::string pooling_type = ctx->Attrs().Get<std::string>("poolingType");
+  std::string pooling_type = ctx->Attrs().Get<std::string>("pooling_type");
   std::vector<int> ksize = ctx->Attrs().Get<std::vector<int>>("ksize");
   std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
   std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
@@ -37,7 +37,7 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
   PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
                  "Pooling intput should be 4-D or 5-D tensor.");
 
-  if (ctx->Attrs().Get<bool>("globalPooling")) {
+  if (ctx->Attrs().Get<bool>("global_pooling")) {
     ksize.resize(static_cast<size_t>(in_x_dims.size()) - 2);
     for (size_t i = 0; i < ksize.size(); ++i) {
       paddings[i] = 0;
@@ -83,20 +83,20 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
             "H is the height of the feature, "
             "and W is the width of the feature.");
 
-  AddAttr<std::string>("poolingType",
+  AddAttr<std::string>("pooling_type",
                        "(string), pooling type, can be \"max\" for max-pooling "
                        "and \"avg\" for average-pooling.")
       .InEnum({"max", "avg"});
   AddAttr<std::vector<int>>("ksize",
                             "(vector<int>) The pooling window "
                             "size(height, width) of the pooling operator. "
-                            "If globalPooling = true, ksize and paddings will "
+                            "If global_pooling = true, ksize and paddings will "
                             "be ignored.");  // TODO(Chengduo): Add checker.
                                              // (Currently,
   // TypedAttrChecker don't support vector type.)
-  AddAttr<bool>("globalPooling",
+  AddAttr<bool>("global_pooling",
                 "(bool, default false) Whether to use the global pooling. "
-                "If globalPooling = true, ksize and paddings will be ignored.")
+                "If global_pooling = true, ksize and paddings will be ignored.")
       .SetDefault(false);
   AddAttr<std::vector<int>>("strides",
                             "(vector<int>, default {1, 1}), strides(height, "
@@ -107,7 +107,7 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
       "paddings",
       "(vector<int>, defalut {0,0}), paddings(height, width) of pooling "
       "operator."
-      "If globalPooling = true, paddings and ksize will be ignored.")
+      "If global_pooling = true, paddings and ksize will be ignored.")
       .SetDefault({0, 0});  // TODO(Chengduo): Add checker. (Currently,
   // TypedAttrChecker don't support vector type.)
 
@@ -115,7 +115,7 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
 Pool2d Operator.
 
 The pooling2d operation calculates the output based on
-the input, poolingType and ksize, strides, paddings parameters.
+the input, pooling_type and ksize, strides, paddings parameters.
 Input(X) and output(Out) are in NCHW format, where N is batch size, C is the
 number of channels, H is the height of the feature, and W is the width of the feature.
 Parameters(ksize, strides, paddings) are two elements.
@@ -152,7 +152,7 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
             "the number of channels, and D, H and W is the depth, height and "
             "width of the feature, respectively.");
 
-  AddAttr<std::string>("poolingType",
+  AddAttr<std::string>("pooling_type",
                        "(string) Pooling type, can be \"max\" for max-pooling "
                        "and \"avg\" for average-pooling.")
       .InEnum({"max", "avg"});
@@ -160,13 +160,14 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
       "ksize",
       "(vector<int>) The pooling window size(depth, height, "
       "width) of pooling operator. "
-      "If globalPooling = true, ksize and paddings will "
+      "If global_pooling = true, ksize and paddings will "
       "be ignored.");  // TODO(Chengduo): Add checker.
                        // (Currently,
   // TypedAttrChecker don't support vector type.)
-  AddAttr<bool>("globalPooling",
-                "(bool, default false) Whether to use the global pooling. "
-                "If globalPooling = true, ksize and paddings wille be ignored.")
+  AddAttr<bool>(
+      "global_pooling",
+      "(bool, default false) Whether to use the global pooling. "
+      "If global_pooling = true, ksize and paddings wille be ignored.")
       .SetDefault(false);
   AddAttr<std::vector<int>>(
       "strides",
@@ -178,7 +179,7 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
       "paddings",
       "(vector<int>, defalut {0,0,0}), paddings(depth, height, "
       "width) of pooling operator. "
-      "If globalPooling = true, ksize and paddings will be ignored.")
+      "If global_pooling = true, ksize and paddings will be ignored.")
       .SetDefault({0, 0, 0});  // TODO(Chengduo): Add checker. (Currently,
                                // TypedAttrChecker don't support vector type.)
 
@@ -186,7 +187,7 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
 Pool3d Operator.
 
 The pooling3d operation calculates the output based on
-the input, poolingType, ksize, strides, and paddings parameters.
+the input, pooling_type, ksize, strides, and paddings parameters.
 Input(X) and output(Out) are in NCDHW format, where N is batch
 size, C is the number of channels, and D, H and W are the depth, height and
 width of the feature, respectively. Parameters(ksize, strides, paddings) 

paddle/operators/pool_op.h

@@ -57,11 +57,11 @@ class PoolKernel : public framework::OpKernel<T> {
     const Tensor* in_x = context.Input<Tensor>("X");
     Tensor* out = context.Output<Tensor>("Out");
 
-    std::string pooling_type = context.Attr<std::string>("poolingType");
+    std::string pooling_type = context.Attr<std::string>("pooling_type");
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
-    if (context.Attr<bool>("globalPooling")) {
+    if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
@@ -119,12 +119,12 @@ class PoolGradKernel : public framework::OpKernel<T> {
         context.Input<Tensor>(framework::GradVarName("Out"));
     Tensor* in_x_grad = context.Output<Tensor>(framework::GradVarName("X"));
 
-    std::string pooling_type = context.Attr<std::string>("poolingType");
+    std::string pooling_type = context.Attr<std::string>("pooling_type");
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
 
-    if (context.Attr<bool>("globalPooling")) {
+    if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(in_x->dims()[i + 2]);

paddle/operators/pool_with_index_op.cc

@@ -44,7 +44,7 @@ class MaxPoolWithIndexOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
                    "Pooling intput should be 4-D or 5-D tensor.");
 
-    if (ctx->Attrs().Get<bool>("globalPooling")) {
+    if (ctx->Attrs().Get<bool>("global_pooling")) {
       ksize.resize(static_cast<size_t>(in_x_dims.size()) - 2);
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
@@ -110,14 +110,14 @@ class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::vector<int>>("ksize",
                               "(vector<int>) The pooling window size(height, "
                               "width) of pooling operator. "
-                              "If globalPooling = true, ksize and paddings "
+                              "If global_pooling = true, ksize and paddings "
                               "will be ignored.");  // TODO(Chengduo): Add
                                                     // checker. (Currently,
     // TypedAttrChecker don't support vector type.)
     AddAttr<bool>(
-        "globalPooling",
+        "global_pooling",
         "(bool, default false) Whether to use the global pooling. "
-        "If globalPooling = true, ksize and paddings will be ignored.")
+        "If global_pooling = true, ksize and paddings will be ignored.")
         .SetDefault(false);
     AddAttr<std::vector<int>>("strides",
                               "(vector<int>, default {1, 1}), strides(height, "
@@ -128,7 +128,7 @@ class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
         "paddings",
         "(vector<int>, defalut {0, 0}), paddings(height, width) of pooling "
         "operator. "
-        "If globalPooling = true, paddings and will be ignored.")
+        "If global_pooling = true, paddings and will be ignored.")
         .SetDefault({0, 0});  // TODO(Chengduo): Add checker. (Currently,
     // TypedAttrChecker don't support vector type.)
 
@@ -188,14 +188,14 @@ class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::vector<int>>("ksize",
                               "(vector<int>) The pooling window size(depth, "
                               "height, width) of pooling operator. "
-                              "If globalPooling = true, ksize and paddings "
+                              "If global_pooling = true, ksize and paddings "
                               "will be ignored.");  // TODO(Chengduo): Add
                                                     // checker. (Currently,
     // TypedAttrChecker don't support vector type.)
     AddAttr<bool>(
-        "globalPooling",
+        "global_pooling",
         "(bool, default false) Whether to use the global pooling. "
-        "If globalPooling = true, ksize and paddings will be ignored.")
+        "If global_pooling = true, ksize and paddings will be ignored.")
         .SetDefault(false);
     AddAttr<std::vector<int>>("strides",
                               "(vector<int>, default {1,1,1}), strides(depth, "
@@ -206,7 +206,7 @@ class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
         "paddings",
         "(vector, defalut {0,0,0}), paddings(depth, "
         "height, width) of pooling operator. "
-        "If globalPooling = true, paddings and ksize will be ignored.")
+        "If global_pooling = true, paddings and ksize will be ignored.")
         .SetDefault({0, 0, 0});  // TODO(Chengduo): Add checker. (Currently,
     // TypedAttrChecker don't support vector type.)
 

paddle/operators/pool_with_index_op.h

@@ -35,7 +35,7 @@ class MaxPoolWithIndexKernel : public framework::OpKernel<T> {
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
-    if (context.Attr<bool>("globalPooling")) {
+    if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
@@ -72,7 +72,7 @@ class MaxPoolWithIndexGradKernel : public framework::OpKernel<T> {
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
-    if (context.Attr<bool>("globalPooling")) {
+    if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(in_x_grad->dims()[i + 2]);

paddle/operators/shrink_rnn_memory_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#include "paddle/framework/lod_rank_table.h"
+#include "paddle/operators/array_operator.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+class ShrinkRNNMemoryOp : public ArrayOp {
+ public:
+  ShrinkRNNMemoryOp(const std::string &type,
+                    const framework::VariableNameMap &inputs,
+                    const framework::VariableNameMap &outputs,
+                    const framework::AttributeMap &attrs)
+      : ArrayOp(type, inputs, outputs, attrs) {}
+
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto *x_var = scope.FindVar(Input("X"));
+    PADDLE_ENFORCE(x_var != nullptr, "Input X must be set");
+    auto &x_tensor = x_var->Get<framework::LoDTensor>();
+    size_t offset = this->GetOffset(scope, dev_ctx);
+    auto *rank_table_var = scope.FindVar(Input("RankTable"));
+    PADDLE_ENFORCE(rank_table_var != nullptr, "RankTable must be set");
+    auto &rank_table = rank_table_var->Get<framework::LoDRankTable>();
+
+    auto &rank_items = rank_table.items();
+    int dst_num_rows =
+        std::lower_bound(rank_items.begin(), rank_items.end(), offset,
+                         [](const framework::LoDRankTable::TableItem &a,
+                            size_t b) { return a.length > b; }) -
+        rank_items.begin();
+
+    auto *out_var = scope.FindVar(Output("Out"));
+    PADDLE_ENFORCE(out_var != nullptr, "Output Out must be set");
+    auto &out_tensor = *out_var->GetMutable<framework::LoDTensor>();
+    if (dst_num_rows != 0) {
+      out_tensor.ShareDataWith(x_tensor.Slice(0, dst_num_rows));
+    }
+  }
+};
+
+class ShrinkRNNMemoryOpProtoMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ShrinkRNNMemoryOpProtoMaker(framework::OpProto *proto,
+                              framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "");
+    AddInput("RankTable", "");
+    AddInput("I", "");
+    AddOutput("Out", "");
+    AddComment("");
+  }
+};
+
+class ShrinkRNNMemoryInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    PADDLE_ENFORCE(context->HasInput("X"));
+    PADDLE_ENFORCE(context->HasInput("I"));
+    PADDLE_ENFORCE(context->HasInput("RankTable"));
+    context->SetOutputDim("Out", context->GetInputDim("X"));
+  }
+};
+
+class ShrinkRNNMemoryGradOp : public ArrayOp {
+ public:
+  ShrinkRNNMemoryGradOp(const std::string &type,
+                        const framework::VariableNameMap &inputs,
+                        const framework::VariableNameMap &outputs,
+                        const framework::AttributeMap &attrs)
+      : ArrayOp(type, inputs, outputs, attrs) {}
+
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto *dout_var = scope.FindVar(Input(framework::GradVarName("Out")));
+    auto *dx_var = scope.FindVar(Output(framework::GradVarName("X")));
+    PADDLE_ENFORCE(dx_var != nullptr, "Input Gradient should not be nullptr");
+    auto *x_var = scope.FindVar(Input("X"));
+    PADDLE_ENFORCE(x_var != nullptr);
+
+    auto &x_tensor = x_var->Get<framework::LoDTensor>();
+    auto &dx_tensor = *dx_var->GetMutable<framework::LoDTensor>();
+    dx_tensor.Resize(x_tensor.dims());
+    dx_tensor.mutable_data(x_tensor.place(), x_tensor.type());
+
+    if (dout_var == nullptr) {  // dx_tensor fill zero
+      math::set_constant(dev_ctx, &dx_tensor, 0.0f);
+    } else {
+      auto &dout_tensor = dout_var->Get<framework::LoDTensor>();
+      auto height = dout_tensor.dims()[0];
+      dx_tensor.Slice(0, static_cast<int>(height))
+          .CopyFrom(dout_tensor, dout_tensor.place(), dev_ctx);
+      if (dx_tensor.dims()[0] < height) {
+        auto rest_tensor = dx_tensor.Slice(
+            static_cast<int>(height), static_cast<int>(dout_tensor.dims()[0]));
+        math::set_constant(dev_ctx, &rest_tensor, 0.0f);
+      }
+    }
+  }
+};
+
+class ShrinkRNNMemoryGradInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    PADDLE_ENFORCE(context->HasInput("X"));
+    PADDLE_ENFORCE(context->HasOutput(framework::GradVarName("X")));
+    context->SetOutputDim(framework::GradVarName("X"),
+                          context->GetInputDim("X"));
+  }
+};
+
+class ShrinkRNNGradOpMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *op = new framework::OpDescBind();
+    op->SetType("shrink_rnn_memory_grad");
+    op->SetInput("X", Input("X"));
+    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDescBind>(op);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(shrink_rnn_memory, ops::ShrinkRNNMemoryOp,
+                  ops::ShrinkRNNMemoryInferShape,
+                  ops::ShrinkRNNMemoryOpProtoMaker, ops::ShrinkRNNGradOpMaker);
+REGISTER_OPERATOR(shrink_rnn_memory_grad, ops::ShrinkRNNMemoryGradOp,
+                  ops::ShrinkRNNMemoryGradInferShape);

paddle/operators/tensor_array_read_write_op.cc

@@ -11,48 +11,18 @@
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License. */
-#include "paddle/framework/lod_tensor_array.h"
-#include "paddle/framework/op_registry.h"
+#include "paddle/operators/array_operator.h"
 
 namespace paddle {
 namespace operators {
-class ArrayOpBase : public framework::OperatorBase {
- public:
-  ArrayOpBase(const std::string &type, const framework::VariableNameMap &inputs,
-              const framework::VariableNameMap &outputs,
-              const framework::AttributeMap &attrs)
-      : OperatorBase(type, inputs, outputs, attrs) {}
-  void Run(const framework::Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {}
-
- protected:
-  size_t GetOffset(const framework::Scope &scope,
-                   const platform::DeviceContext &dev_ctx) const {
-    auto *i = scope.FindVar(Input("I"));
-    PADDLE_ENFORCE(i != nullptr, "I must be set");
-    auto &i_tensor = i->Get<framework::LoDTensor>();
-    PADDLE_ENFORCE_EQ(i_tensor.numel(), 1);
-    size_t offset;
-    if (platform::is_gpu_place(i_tensor.place())) {
-      // FIXME: Avoid copy from GPU to CPU
-      framework::Tensor t;
-      t.CopyFrom(i_tensor, platform::CPUPlace(), dev_ctx);
-      dev_ctx.Wait();
-      offset = static_cast<size_t>(*t.data<int64_t>());
-    } else {
-      offset = static_cast<size_t>(*i_tensor.data<int64_t>());
-    }
-    return offset;
-  }
-};
 
-class WriteToArrayOp : public ArrayOpBase {
+class WriteToArrayOp : public ArrayOp {
  public:
   WriteToArrayOp(const std::string &type,
                  const framework::VariableNameMap &inputs,
                  const framework::VariableNameMap &outputs,
                  const framework::AttributeMap &attrs)
-      : ArrayOpBase(type, inputs, outputs, attrs) {}
+      : ArrayOp(type, inputs, outputs, attrs) {}
 
   void Run(const framework::Scope &scope,
            const platform::DeviceContext &dev_ctx) const override {
@@ -122,13 +92,13 @@ class WriteToArrayInferVarType : public framework::VarTypeInference {
   }
 };
 
-class ReadFromArrayOp : public ArrayOpBase {
+class ReadFromArrayOp : public ArrayOp {
  public:
   ReadFromArrayOp(const std::string &type,
                   const framework::VariableNameMap &inputs,
                   const framework::VariableNameMap &outputs,
                   const framework::AttributeMap &attrs)
-      : ArrayOpBase(type, inputs, outputs, attrs) {}
+      : ArrayOp(type, inputs, outputs, attrs) {}
   void Run(const framework::Scope &scope,
            const platform::DeviceContext &dev_ctx) const override {
     auto *x = scope.FindVar(Input("X"));

paddle/platform/transform.h

@@ -49,8 +49,6 @@ struct Transform<platform::CPUPlace> {
   template <typename InputIter, typename OutputIter, typename UnaryOperation>
   void operator()(const DeviceContext& context, InputIter first, InputIter last,
                   OutputIter result, UnaryOperation op) {
-    auto place = context.GetPlace();
-    PADDLE_ENFORCE(is_cpu_place(place), "It must use CPU place.");
     std::transform(first, last, result, op);
   }
 
@@ -59,8 +57,6 @@ struct Transform<platform::CPUPlace> {
   void operator()(const DeviceContext& context, InputIter1 first1,
                   InputIter1 last1, InputIter2 first2, OutputIter result,
                   BinaryOperation op) {
-    auto place = context.GetPlace();
-    PADDLE_ENFORCE(is_cpu_place(place), "It must use CPU place.");
     std::transform(first1, last1, first2, result, op);
   }
 };

proto/ModelConfig.proto

@@ -321,6 +321,11 @@ message ClipConfig {
   required double max = 2;
 }
 
+message ScaleSubRegionConfig {
+  required ImageConfig image_conf = 1;
+  required float value = 2;
+}
+
 message LayerInputConfig {
   required string input_layer_name = 1;
   optional string input_parameter_name = 2;
@@ -342,6 +347,7 @@ message LayerInputConfig {
   optional MultiBoxLossConfig multibox_loss_conf = 16;
   optional DetectionOutputConfig detection_output_conf = 17;
   optional ClipConfig clip_conf = 18;
+  optional ScaleSubRegionConfig scale_sub_region_conf = 19;
 }
 
 message LayerConfig {

python/paddle/trainer/config_parser.py

@@ -3801,6 +3801,25 @@ class SwitchOrderLayer(LayerBase):
         self.config.reshape_conf.width_axis.extend(reshape['width'])
 
 
+@config_layer('scale_sub_region')
+class ScaleSubRegionLayer(LayerBase):
+    def __init__(self, name, inputs, value, **xargs):
+        super(ScaleSubRegionLayer, self).__init__(
+            name, 'scale_sub_region', 0, inputs=inputs, **xargs)
+        scale_sub_region_conf = self.config.inputs[0].scale_sub_region_conf
+        scale_sub_region_conf.value = value
+
+        # get channel, width and height from input_0 layer
+        input_layer = self.get_input_layer(0)
+        image_conf = scale_sub_region_conf.image_conf
+        image_conf.img_size = input_layer.width
+        image_conf.img_size_y = input_layer.height
+        image_conf.channels = input_layer.size / (input_layer.width *
+                                                  input_layer.height)
+        self.set_cnn_layer(name, image_conf.img_size_y, image_conf.img_size,
+                           image_conf.channels)
+
+
 # Deprecated, use a new layer specific class instead
 @config_func
 def Layer(name, type, **xargs):

python/paddle/trainer_config_helpers/layers.py

@@ -144,6 +144,7 @@ __all__ = [
     'img_conv3d_layer',
     'resize_layer',
     'sub_seq_layer',
+    'scale_sub_region_layer',
 ]
 
 
@@ -255,6 +256,8 @@ class LayerType(object):
     RESIZE = 'resize'
     SUB_SEQ_LAYER = 'subseq'
 
+    SCALE_SUB_REGION_LAYER = 'scale_sub_region'
+
     @staticmethod
     def is_layer_type(type_name):
         """
@@ -786,10 +789,9 @@ class MixedLayerType(LayerOutput):
         :type size: int
         :param act: Activation type.
         :type act: BaseActivation
-        :param bias_attr: The Bias Attribute. If the parameter is set to
-                          False or something not type of ParameterAttribute,
-                          no bias is defined. If the parameter is set to
-                          True, the bias is initialized to zero.
+        :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                          whose type is not ParameterAttribute, no bias is defined. If the
+                          parameter is set to True, the bias is initialized to zero.
         :type bias_attr: ParameterAttribute | None | bool | Any
         :param layer_attr: Extra Layer Attribute.
         :type layer_attr: ExtraLayerAttribute or None
@@ -886,10 +888,9 @@ def mixed_layer(size=0,
                   then this function will just return layer's name.
     :param act: Activation Type. LinearActivation is the default.
     :type act: BaseActivation
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: The extra layer config. Default is None.
     :type layer_attr: ExtraLayerAttribute
@@ -1031,10 +1032,9 @@ def fc_layer(input,
     :type act: BaseActivation
     :param param_attr: The Parameter Attribute|list.
     :type param_attr: ParameterAttribute
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer config.
     :type layer_attr: ExtraLayerAttribute | None
@@ -1387,10 +1387,9 @@ def pooling_layer(input,
     :type pooling_type: BasePoolingType | None
     :param stride: The step size between successive pooling regions.
     :type stride: Int
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: The Extra Attributes for layer, such as dropout.
     :type layer_attr: ExtraLayerAttribute | None
@@ -1488,10 +1487,9 @@ def lstmemory(input,
     :type gate_act: BaseActivation
     :param state_act: state activation type, TanhActivation by default.
     :type state_act: BaseActivation
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: Parameter Attribute.
     :type param_attr: ParameterAttribute | None | False
@@ -1614,10 +1612,9 @@ def grumemory(input,
                      This activation affects the :math:`z_t` and :math:`r_t`. It is the
                      :math:`\\sigma` in the above formula.
     :type gate_act: BaseActivation
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: Parameter Attribute.
     :type param_attr: ParameterAttribute | None | False
@@ -1814,10 +1811,9 @@ def expand_layer(input,
     :type expand_as: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param expand_level: whether input layer is timestep(default) or sequence.
     :type expand_level: ExpandLevel
@@ -1936,10 +1932,9 @@ def seq_reshape_layer(input,
     :type act: BaseActivation
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -2323,10 +2318,9 @@ def hsigmoid(input,
     :type num_classes: int | None
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: Parameter Attribute. None means default parameter.
     :type param_attr: ParameterAttribute | None
@@ -2466,10 +2460,9 @@ def img_conv_layer(input,
     :type dilation: int | tuple | list
     :param dilation_y: The y dimension of the dilation.
     :type dilation_y: int
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param num_channels: number of input channels. If None will be set
                         automatically from previous output.
@@ -3216,10 +3209,9 @@ def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None):
     :type input: LayerOutput | list | tuple
     :param act: Activation Type. LinearActivation is the default.
     :type act: BaseActivation
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer attribute.
     :type layer_attr: ExtraLayerAttribute
@@ -3372,10 +3364,9 @@ def seq_concat_layer(a, b, act=None, name=None, layer_attr=None,
     :type act: BaseActivation
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -3555,10 +3546,9 @@ def lstm_step_layer(input,
     :type gate_act: BaseActivation
     :param state_act: State Activation Type. TanhActivation is the default.
     :type state_act: BaseActivation
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: layer's extra attribute.
     :type layer_attr: ExtraLayerAttribute
@@ -3614,10 +3604,9 @@ def gru_step_layer(input,
     :param name: The name of this layer. It is optional.
     :param gate_act: Activation type of this layer's two gates. Default is Sigmoid.
     :type gate_act: BaseActivation
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: the parameter_attribute for transforming the output_mem
                        from previous step.
@@ -3677,10 +3666,9 @@ def gru_step_naive_layer(input,
     :type act: BaseActivation
     :param gate_act: Activation type of this layer's two gates. Default is Sigmoid.
     :type gate_act: BaseActivation
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr:
     :param layer_attr:
@@ -3810,10 +3798,9 @@ def recurrent_layer(input,
     :type input: LayerOutput
     :param act: Activation type. TanhActivation is the default.
     :type act: BaseActivation
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: parameter attribute.
     :type param_attr: ParameterAttribute
@@ -4803,10 +4790,9 @@ def tensor_layer(a,
     :type act: BaseActivation
     :param param_attr: The Parameter Attribute.
     :type param_attr: ParameterAttribute
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer config.
     :type layer_attr: ExtraLayerAttribute | None
@@ -4868,10 +4854,9 @@ def selective_fc_layer(input,
     :type act: BaseActivation
     :param param_attr: The Parameter Attribute.
     :type param_attr: ParameterAttribute
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer config.
     :type layer_attr: ExtraLayerAttribute | None
@@ -5494,7 +5479,11 @@ def crf_decoding_layer(input,
     return LayerOutput(name, LayerType.CRF_DECODING_LAYER, parents, size=1)
 
 
-@wrap_act_default(act=SigmoidActivation())
+"""
+Following are cost Layers.
+"""
+
+
 @wrap_bias_attr_default(has_bias=True)
 @wrap_param_attr_default()
 @wrap_name_default()
@@ -5502,7 +5491,6 @@ def crf_decoding_layer(input,
 def nce_layer(input,
               label,
               num_classes=None,
-              act=None,
               param_attr=None,
               weight=None,
               num_neg_samples=10,
@@ -5511,9 +5499,12 @@ def nce_layer(input,
               bias_attr=None,
               layer_attr=None):
     """
-    Noise-contrastive estimation.
-    Implements the method in the following paper:
-    A fast and simple algorithm for training neural probabilistic language models.
+    Noise-contrastive estimation. This layer implements the method in the
+    following paper:
+
+    Reference:
+        A fast and simple algorithm for training neural probabilistic language
+        models. https://www.cs.toronto.edu/~amnih/papers/ncelm.pdf
 
     The example usage is:
 
@@ -5525,32 +5516,37 @@ def nce_layer(input,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layers. It could be a LayerOutput of list/tuple of LayerOutput.
+    :param input: The input layers. It should be a LayerOutput or a list/tuple
+                  of LayerOutput.
     :type input: LayerOutput | list | tuple | collections.Sequence
-    :param label: label layer
+    :param label: The ground truth.
     :type label: LayerOutput
-    :param weight: weight layer, can be None(default)
+    :param weight: The weight layer defines a weight for each sample in the
+                   mini-batch. The default value is None.
     :type weight: LayerOutput
-    :param num_classes: number of classes.
+    :param num_classes: The class number.
     :type num_classes: int
-    :param act: Activation type. SigmoidActivation is the default.
-    :type act: BaseActivation
-    :param param_attr: The Parameter Attribute|list.
-    :type param_attr: ParameterAttribute
-    :param num_neg_samples: number of negative samples. Default is 10.
+    :param param_attr: The parameter attributes.
+    :type param_attr: ParameterAttribute|list
+    :param num_neg_samples: The number of sampled negative labels. The default
+                            value is 10.
     :type num_neg_samples: int
-    :param neg_distribution: The distribution for generating the random negative labels.
-                             A uniform distribution will be used if not provided.
-                             If not None, its length must be equal to num_classes.
+    :param neg_distribution: The discrete noisy distribution over the output
+                             space from which num_neg_samples negative labels
+                             are sampled. If this parameter is not set, a
+                             uniform distribution will be used. A user defined
+                             distribution is a list whose length must be equal
+                             to the num_classes. Each member of the list defines
+                             the probability of a class given input x.
     :type neg_distribution: list | tuple | collections.Sequence | None
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The attribute for bias. If this parameter is set False or
+                      any object whose type is not ParameterAttribute, no bias
+                      is added. If this parameter is set True, the bias is
+                      initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
-    :return: layer name.
+    :return: The LayerOutput object.
     :rtype: LayerOutput
     """
     if isinstance(input, LayerOutput):
@@ -5573,8 +5569,6 @@ def nce_layer(input,
         assert isinstance(neg_distribution, collections.Sequence)
         assert len(neg_distribution) == num_classes
         assert abs(sum(neg_distribution) - 1.0) < 1e-5
-    if not isinstance(act, BaseActivation):
-        raise TypeError()
 
     ipts_for_layer = []
     parents = []
@@ -5596,7 +5590,7 @@ def nce_layer(input,
         type=LayerType.NCE_LAYER,
         num_classes=num_classes,
         neg_sampling_dist=neg_distribution,
-        active_type=act.name,
+        active_type=SigmoidActivation().name,
         num_neg_samples=num_neg_samples,
         inputs=ipts_for_layer,
         bias=ParamAttr.to_bias(bias_attr),
@@ -5606,12 +5600,7 @@ def nce_layer(input,
         LayerType.NCE_LAYER,
         parents=parents,
         size=l.config.size,
-        activation=act)
-
-
-"""
-following are cost Layers.
-"""
+        activation=SigmoidActivation())
 
 
 @wrap_name_default()
@@ -5770,20 +5759,21 @@ def cross_entropy(input,
     :param input: The first input layer.
     :type input: LayerOutput.
     :param label: The input label.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None | basestring.
-    :param coeff: The cost is multiplied with coeff.
-                  The coefficient affects the gradient in the backward.
-    :type coeff: float.
+    :type name: basestring
+    :param coeff: The weight of the gradient in the back propagation.
+                  1.0 is the default.
+    :type coeff: float
     :param weight: The cost of each sample is multiplied with each weight.
                    The weight should be a layer with size=1. Note that gradient
                    will not be calculated for weight.
     :type weight: LayerOutout
-    :param layer_attr: Extra Layer Attribute.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
-    :rtype: LayerOutput.
+    :rtype: LayerOutput
     """
 
     ipts, parents = __cost_input__(input, label, weight)
@@ -5816,19 +5806,21 @@ def cross_entropy_with_selfnorm(input,
                                           label=label_layer)
 
     :param input: The first input layer.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param label: The input label.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None | basestring.
-    :param coeff: The coefficient affects the gradient in the backward.
-    :type coeff: float.
+    :type name: basestring
+    :param coeff: The weight of the gradient in the back propagation.
+                  1.0 is the default.
+    :type coeff: float
     :param softmax_selfnorm_alpha: The scale factor affects the cost.
-    :type softmax_selfnorm_alpha: float.
-    :param layer_attr: Extra Layer Attribute.
+    :type softmax_selfnorm_alpha: float
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
-    :rtype: LayerOutput.
+    :rtype: LayerOutput
     """
     Layer(
         name=name,
@@ -5849,7 +5841,7 @@ def cross_entropy_with_selfnorm(input,
 @layer_support()
 def sum_cost(input, name=None, layer_attr=None):
     """
-    A loss layer which calculate the sum of the input as loss
+    A loss layer which calculates the sum of the input as loss.
 
     The example usage is:
 
@@ -5858,10 +5850,11 @@ def sum_cost(input, name=None, layer_attr=None):
        cost = sum_cost(input=input_layer)
 
     :param input: The input of this layer.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None | basestring.
-    :param layer_attr: Extra Layer Attribute.
+    :type name: basestring
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
     :rtype: LayerOutput.
@@ -5901,16 +5894,18 @@ def huber_regression_cost(input,
        cost = huber_regression_cost(input=input_layer, label=label_layer)
 
     :param input: The first input layer.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param label: The input label.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None | basestring.
+    :type name: basestring
     :param delta: The difference between the observed and predicted values.
-    :type delta: float.
-    :param coeff: The coefficient affects the gradient in the backward.
-    :type coeff: float.
-    :param layer_attr: Extra Layer Attribute.
+    :type delta: float
+    :param coeff: The weight of the gradient in the back propagation.
+                  1.0 is the default.
+    :type coeff: float
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
     :rtype: LayerOutput.
@@ -5951,17 +5946,19 @@ def huber_classification_cost(input,
        cost = huber_classification_cost(input=input_layer, label=label_layer)
 
     :param input: The first input layer.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param label: The input label.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None | basestring.
-    :param coeff: The coefficient affects the gradient in the backward.
-    :type coeff: float.
-    :param layer_attr: Extra Layer Attribute.
+    :type name: basestring
+    :param coeff: The weight of the gradient in the back propagation.
+                  1.0 is the default.
+    :type coeff: float
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
-    :rtype: LayerOutput.
+    :rtype: LayerOutput
     """
     assert isinstance(input, LayerOutput)
     if input.size is not None:
@@ -5998,10 +5995,12 @@ def multi_binary_label_cross_entropy(input,
     :param label: The input label.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None | basestring
-    :param coeff: The coefficient affects the gradient in the backward.
+    :type name: basestring
+    :param coeff: The weight of the gradient in the back propagation.
+                  1.0 is the default.
     :type coeff: float
-    :param layer_attr: Extra Layer Attribute.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -6104,7 +6103,7 @@ def cross_entropy_over_beam(input, name=None):
 
     :param input: Input beams for this layer.
     :type input: BeamInput
-    :param name: The name of this layer.
+    :param name: The name of this layer. It is optional.
     :type name: basestring
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -6139,7 +6138,7 @@ def cross_entropy_over_beam(input, name=None):
 def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None):
     """
     This is a L1 loss but more smooth. It requires that the
-    size of input and label are equal. The formula is as follows,
+    sizes of input and label are equal. The formula is as follows,
 
     .. math::
 
@@ -6151,8 +6150,9 @@ def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None):
 
         smooth_{L1}(x) = \\begin{cases} 0.5x^2& \\text{if}  \\ |x| < 1 \\\\ |x|-0.5& \\text{otherwise} \end{cases}
 
-    More details can be found by referring to `Fast R-CNN
-    <https://arxiv.org/pdf/1504.08083v2.pdf>`_
+    Reference:
+        Fast R-CNN
+        https://arxiv.org/pdf/1504.08083v2.pdf
 
     The example usage is:
 
@@ -6166,10 +6166,12 @@ def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None):
     :param label: The input label.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None | basestring
-    :param coeff: The coefficient affects the gradient in the backward.
+    :type name: basestring
+    :param coeff: The weight of the gradient in the back propagation.
+                  1.0 is the default.
     :type coeff: float
-    :param layer_attr: Extra Layer Attribute.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -6191,12 +6193,12 @@ def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None):
 @wrap_name_default()
 def multiplex_layer(input, name=None, layer_attr=None):
     """
-    This layer multiplex multiple layers according to the index,
-    which is provided by the first input layer.
-    inputs[0]: the index of the layer to output of size batchSize.
+    This layer multiplex multiple layers according to the indexes,
+    which are provided by the first input layer.
+    inputs[0]: the indexes of the layers to form the output of size batchSize.
     inputs[1:N]; the candidate output data.
-    For each index i from 0 to batchSize -1, the output is the i-th row of the
-    (index[i] + 1)-th layer.
+    For each index i from 0 to batchSize - 1, the i-th row of the output is the
+    the same to the i-th row of the (index[i] + 1)-th layer.
 
     For each i-th row of output:
     .. math::
@@ -6215,7 +6217,8 @@ def multiplex_layer(input, name=None, layer_attr=None):
     :type input: list of LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param layer_attr: extra layer attributes.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -6319,14 +6322,14 @@ def row_conv_layer(input,
     :type context_len: int
     :param act: Activation Type. LinearActivation is the default.
     :type act: BaseActivation
-    :param param_attr: The Parameter Attribute. If None, the parameter will be
-                       initialized smartly. It's better to set it by yourself.
+    :param param_attr: The parameter attribute. See ParameterAttribute for
+                       details.
     :type param_attr: ParameterAttribute
-    :param layer_attr: Extra Layer config.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
-
     """
     assert isinstance(input, LayerOutput)
     assert context_len > 0, "the context_len must be greatet than 0."
@@ -6351,7 +6354,7 @@ def prelu_layer(input,
                 param_attr=None,
                 layer_attr=None):
     """
-    The Parameter Relu activation that actives outputs with a learnable weight.
+    The Parametric Relu activation that actives outputs with a learnable weight.
 
     Reference:
         Delving Deep into Rectifiers: Surpassing Human-Level Performance on
@@ -6371,16 +6374,17 @@ def prelu_layer(input,
     :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param partial_sum: this parameter makes a group of inputs share a same weight.
+    :param partial_sum: this parameter makes a group of inputs share the same weight.
 
         - partial_sum = 1, indicates the element-wise activation: each element has a weight.
-        - partial_sum = number of elements in one channel, indicates the channel-wise activation, elements in a channel share a same weight.
-        - partial_sum = number of outputs, indicates all elements share a same weight.
+        - partial_sum = number of elements in one channel, indicates the channel-wise activation, elements in a channel share the same weight.
+        - partial_sum = number of outputs, indicates all elements share the same weight.
 
     :type partial_sum: int
     :param param_attr: The parameter attribute. See ParameterAttribute for details.
-    :type param_attr: ParameterAttribute | None
-    :param layer_attr: Extra layer configurations. Default is None.
+    :type param_attr: ParameterAttribute
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -6436,34 +6440,34 @@ def gated_unit_layer(input,
 
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param size: output size of the gated unit.
+    :param size: The dimension of this layer's output.
     :type size: int
-    :param act: Activation type of the projected input. LinearActivation is the default.
+    :param act: Activation type of the projection. LinearActivation is the default.
     :type act: BaseActivation
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param gate_attr: Attributes to tune the gate output, for example, error
-        clipping threshold, dropout and so on. See ExtraLayerAttribute for
-        more details.
+    :param gate_attr: The extra layer attribute of the gate. See ExtraLayerAttribute for
+                      details.
     :type gate_attr: ExtraLayerAttribute | None
-    :param gate_param_attr: Attributes to tune the learnable projected matrix
-        parameter of the gate.
-    :type gate_param_attr: ParameterAttribute | None
-    :param gate_bias_attr: Attributes to tune the learnable bias of the gate.
-    :type gate_bias_attr: ParameterAttribute | None
-    :param inproj_attr: Attributes to the tune the projected input, for
-        example, error clipping threshold, dropout and so on. See
-        ExtraLayerAttribute for more details.
+    :param gate_param_attr: The parameter attribute of the gate. See ParameterAttribute
+                            for details.
+    :type gate_param_attr: ParameterAttribute
+    :param gate_bias_attr: The bias attribute of the gate. If the parameter is set to False or
+                           an object whose type is not ParameterAttribute, no bias is defined.
+                           If the parameter is set to True, the bias is initialized to zero.
+    :type gate_bias_attr: ParameterAttribute | bool | None | Any
+    :param inproj_attr: Extra layer attributes of the projection. See ExtraLayerAttribute for
+                        details.
     :type inproj_attr: ExtraLayerAttribute | None
-    :param inproj_param_attr: Attributes to tune the learnable parameter of
-        the projection of input.
-    :type inproj_param_attr: ParameterAttribute | None
-    :param inproj_bias_attr: Attributes to tune the learnable bias of
-        projection of the input.
-    :type inproj_bias_attr: ParameterAttribute | None
-    :param layer_attr: Attributes to tune the final output of the gated unit,
-        for example, error clipping threshold, dropout and so on. See
-        ExtraLayerAttribute for more details.
+    :param inproj_param_attr: The parameter attribute of the projection. See ParameterAttribute
+                              for details.
+    :type inproj_param_attr: ParameterAttribute
+    :param inproj_bias_attr: The bias attribute of the projection. If the parameter is set to False
+                             or an object whose type is not ParameterAttribute, no bias is defined.
+                             If the parameter is set to True, the bias is initialized to zero.
+    :type inproj_bias_attr: ParameterAttribute | bool | None | Any
+    :param layer_attr: Extra layer attribute of the product. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -6659,9 +6663,9 @@ def clip_layer(input, min, max, name=None):
     :param input: The input of this layer.
     :type input: LayerOutput.
     :param min: The lower threshold for clipping.
-    :type min: double
+    :type min: float
     :param max: The upper threshold for clipping.
-    :type max: double
+    :type max: float
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -6709,7 +6713,6 @@ def seq_slice_layer(input, starts, ends, name=None):
     :type ends: LayerOutput | None
     :return: LayerOutput object.
     :rtype: LayerOutput
-
     """
 
     assert isinstance(input, LayerOutput), (
@@ -6830,20 +6833,21 @@ def img_conv3d_layer(input,
     :param padding: The numbers of padding along three axises. If the parameter is set to
                     one integer, they will be same.
     :type padding: int | tuple | list
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :param num_channels: The number of input channels. If the parameter is not set or
                          set to None,  its actual value will be automatically set to
                          the channels number of the input .
     :type num_channels: int
-    :param param_attr: The parameter attribute of the convolution.
+    :param param_attr: The parameter attribute of the convolution. See ParameterAttribute for
+                       details.
     :type param_attr: ParameterAttribute
     :param shared_biases: Whether biases will be shared between filters or not.
     :type shared_biases: bool
-    :param layer_attr: Extra layer attributes.
+    :param layer_attr: The extra layer attributes. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute
     :param trans: True if it is a convTransLayer, False if it is a convLayer
     :type trans: bool
@@ -6950,12 +6954,12 @@ def scale_shift_layer(input, name=None, param_attr=None, bias_attr=None):
     :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param param_attr: The parameter attribute of scaling.
+    :param param_attr: The parameter attribute of scaling. See ParameterAttribute for
+                      details.
     :type param_attr: ParameterAttribute
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -7013,10 +7017,9 @@ def sub_seq_layer(input, offsets, sizes, act=None, bias_attr=None, name=None):
     :type sizes: LayerOutput
     :param act: Activation type, LinearActivation is the default.
     :type act: BaseActivation.
-    :param bias_attr: The Bias Attribute. If the parameter is set to
-                      False or something not type of ParameterAttribute,
-                      no bias is defined. If the parameter is set to
-                      True, the bias is initialized to zero.
+    :param bias_attr: The bias attribute. If the parameter is set to False or an object
+                      whose type is not ParameterAttribute, no bias is defined. If the
+                      parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -7042,3 +7045,54 @@ def sub_seq_layer(input, offsets, sizes, act=None, bias_attr=None, name=None):
         LayerType.SUB_SEQ_LAYER,
         parents=[input, offsets, sizes],
         size=input.size)
+
+
+@wrap_name_default('scale_sub_region')
+def scale_sub_region_layer(input, indices, value, name=None):
+    """
+    Given an image or feature map with CHW information, scale_sub_region_layer
+    can be used to multiply a real value to values of a sub continuous region.
+    You can provide start and end indices of CHW for each instance.
+    Please notice that all start indices are counting from 1.
+    The shape of indices should be [batch_size, 6] and the layout for each row
+    is [C_Start, C_End, H_Start, H_End, W_Start, W_End].
+
+    .. code-block:: python
+
+        scale_sub_region = scale_sub_region_layer(input=input,
+                                                  indices=indices,
+                                                  value=value)
+
+    :param name: The name of this layer. It is optional.
+    :type name: basestring
+    :param input: The input of this layer which should contains CHW information.
+    :type input: LayerOutput
+    :param indices: Start index and end index for C H W, the input value should
+                    be a 2-D matrix with shape [batch_size, 6].
+    :type indices: LayerOutput.
+    :param value: value to multiply.
+    :type value: float
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+
+    assert isinstance(input, LayerOutput), (
+        'The first input of scale_sub_region_layer, '
+        'must be a PaddlePaddle layer.')
+    assert isinstance(indices, LayerOutput), (
+        'The start and end indices for CHW, must be a PaddlePaddle layer.')
+    assert isinstance(value, float), (
+        'The value to multiply, must be a real value.')
+
+    Layer(
+        name=name,
+        type=LayerType.SCALE_SUB_REGION_LAYER,
+        inputs=[input.name, indices.name],
+        value=value)
+
+    return LayerOutput(
+        name,
+        LayerType.SCALE_SUB_REGION_LAYER,
+        parents=[input, indices],
+        num_filters=input.num_filters,
+        size=input.size)

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -10,6 +10,6 @@ test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_la
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
 test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer
 test_seq_slice_layer test_cross_entropy_over_beam test_pooling3D_layer
-test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer)
+test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer test_scale_sub_region_layer)
 
 export whole_configs=(test_split_datasource)

python/paddle/trainer_config_helpers/tests/configs/protostr/test_scale_sub_region_layer.protostr

+type: "nn"
+layers {
+  name: "data"
+  type: "data"
+  size: 2016
+  active_type: ""
+  height: 48
+  width: 42
+}
+layers {
+  name: "indices"
+  type: "data"
+  size: 6
+  active_type: ""
+}
+layers {
+  name: "__scale_sub_region_0__"
+  type: "scale_sub_region"
+  size: 2016
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    scale_sub_region_conf {
+      image_conf {
+        channels: 1
+        img_size: 42
+        img_size_y: 48
+      }
+      value: 0.0
+    }
+  }
+  inputs {
+    input_layer_name: "indices"
+  }
+  height: 48
+  width: 42
+}
+input_layer_names: "data"
+input_layer_names: "indices"
+output_layer_names: "__scale_sub_region_0__"
+sub_models {
+  name: "root"
+  layer_names: "data"
+  layer_names: "indices"
+  layer_names: "__scale_sub_region_0__"
+  input_layer_names: "data"
+  input_layer_names: "indices"
+  output_layer_names: "__scale_sub_region_0__"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/test_scale_sub_region_layer.py

+from paddle.trainer_config_helpers import *
+
+settings(batch_size=1000, learning_rate=1e-5)
+
+data = data_layer(name='data', size=2016, height=48, width=42)
+indices = data_layer(name='indices', size=6)
+
+scale_sub_region = scale_sub_region_layer(
+    input=data, indices=indices, value=0.0)
+
+outputs(scale_sub_region)

python/paddle/v2/dataset/uci_housing.py

@@ -22,6 +22,7 @@ parse training set and test set into paddle reader creators.
 import numpy as np
 import os
 import paddle.v2.dataset.common
+from paddle.v2.parameters import Parameters
 
 __all__ = ['train', 'test']
 
@@ -34,7 +35,8 @@ feature_names = [
 
 UCI_TRAIN_DATA = None
 UCI_TEST_DATA = None
-
+URL_MODEL = 'https://github.com/PaddlePaddle/book/raw/develop/01.fit_a_line/fit_a_line.tar'
+MD5_MODEL = '52fc3da8ef3937822fcdd87ee05c0c9b'
 
 def feature_range(maximums, minimums):
     import matplotlib
@@ -111,6 +113,13 @@ def test():
     return reader
 
 
+def model():
+    tar_file = paddle.v2.dataset.common.download(URL_MODEL, 'fit_a_line.tar', MD5_MODEL)
+    with open(tar_file, 'r') as f:
+        parameters = Parameters.from_tar(f)
+    return parameters
+
+
 def fetch():
     paddle.v2.dataset.common.download(URL, 'uci_housing', MD5)
 

python/paddle/v2/framework/layers.py

@@ -87,7 +87,8 @@ def data(name,
          type=core.VarDesc.VarType.LOD_TENSOR,
          append_batch_size=True,
          main_program=None,
-         startup_program=None):
+         startup_program=None,
+         stop_gradient=True):
     helper = LayerHelper('data', **locals())
     shape = list(shape)
     for i in xrange(len(shape)):
@@ -101,7 +102,11 @@ def data(name,
         shape = [-1] + shape  # append batch size as -1
 
     return helper.create_global_variable(
-        name=name, shape=shape, dtype=data_type, type=type, stop_gradient=True)
+        name=name,
+        shape=shape,
+        dtype=data_type,
+        type=type,
+        stop_gradient=stop_gradient)
 
 
 def _convert_(name):
@@ -134,9 +139,7 @@ def _create_op_func_(op_type):
     o_name = not_intermediate_outputs[0].name
     intermediate_output_names = [output.name for output in intermediate_outputs]
 
-    def func(**kwargs):
-        helper = LayerHelper(op_type, **kwargs)
-        inputs = dict()
+    def infer_and_check_data_type(op_proto, **kwargs):
         dtype = None
         for ipt in op_proto.inputs:
             name = _convert_(ipt.name)
@@ -153,6 +156,20 @@ def _create_op_func_(op_type):
                 elif dtype != each.data_type:
                     raise ValueError(
                         "operator {0} must input same dtype".format(op_type))
+
+        return dtype
+
+    def func(**kwargs):
+        helper = LayerHelper(op_type, **kwargs)
+
+        dtype = infer_and_check_data_type(op_proto, **kwargs)
+
+        inputs = dict()
+        for ipt in op_proto.inputs:
+            name = _convert_(ipt.name)
+            val = kwargs.pop(name, [])
+            if not isinstance(val, list) and not isinstance(val, tuple):
+                val = [val]
             inputs[ipt.name] = val
 
         outputs = dict()
@@ -178,6 +195,20 @@ _create_op_func_('reshape')
 _create_op_func_('elementwise_add')
 _create_op_func_('sigmoid')
 _create_op_func_('scale')
+_create_op_func_('reshape')
+_create_op_func_('transpose')
+
+
+def fill_constant(data_type, shape, value=None, program=None):
+    helper = LayerHelper('fill_constant', **locals())
+    out = helper.create_tmp_variable(dtype=data_type)
+    helper.append_op(
+        type='fill_constant',
+        outputs={'Out': [out]},
+        attrs={'data_type': data_type,
+               'shape': shape,
+               'value': value})
+    return out
 
 
 def cast(x, data_type, main_program=None):
@@ -414,9 +445,9 @@ def pool2d(input,
         inputs={"X": input},
         outputs={"Out": pool_out},
         attrs={
-            "poolingType": pool_type,
+            "pooling_type": pool_type,
             "ksize": pool_size,
-            "globalPooling": global_pooling,
+            "global_pooling": global_pooling,
             "strides": pool_stride,
             "paddings": pool_padding
         })
@@ -762,6 +793,46 @@ class StaticRNN(object):
             })
 
 
+def lstm(x,
+         c_pre_init,
+         hidden_dim,
+         forget_bias=None,
+         main_program=None,
+         startup_program=None):
+    helper = LayerHelper('lstm_unit', **locals())
+    rnn = StaticRNN()
+    with rnn.step():
+        c_pre = rnn.memory(init=c_pre_init)
+        x_t = rnn.step_input(x)
+
+        before_fc = concat(
+            input=[x_t, c_pre],
+            axis=1,
+            main_program=main_program,
+            startup_program=startup_program)
+        after_fc = fc(input=before_fc,
+                      size=hidden_dim * 4,
+                      main_program=main_program,
+                      startup_program=startup_program)
+
+        data_type = x.data_type
+        c = helper.create_tmp_variable(data_type)
+        h = helper.create_tmp_variable(data_type)
+
+        helper.append_op(
+            type='lstm_unit',
+            inputs={"X": after_fc,
+                    "C_prev": c_pre},
+            outputs={"C": c,
+                     "H": h},
+            attrs={"forget_bias": forget_bias})
+
+        rnn.update_memory(c_pre, c)
+        rnn.output(h)
+
+    return rnn()
+
+
 def lod_rank_table(x, level=0, main_program=None):
     helper = LayerHelper("lod_rank_table", **locals())
     table = helper.create_variable(
@@ -779,7 +850,8 @@ def lod_tensor_to_array(x, table, main_program=None):
     helper = LayerHelper("lod_tensor_to_array", **locals())
     array = helper.create_variable(
         name=unique_name("lod_tensor_to_array"),
-        type=core.VarDesc.VarType.LOD_TENSOR_ARRAY)
+        type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
+        dtype=x.data_type)
     helper.append_op(
         type='lod_tensor_to_array',
         inputs={'X': x,
@@ -826,15 +898,15 @@ def zeros(shape, dtype, main_program=None):
 def increment(x, value=1.0, in_place=True, main_program=None):
     helper = LayerHelper("increment", **locals())
     if in_place:
-        tmp = x
+        out = x
     else:
-        tmp = helper.create_tmp_variable(dtype=x.data_type)
+        out = helper.create_tmp_variable(dtype=x.data_type)
     helper.append_op(
         type='increment',
         inputs={'X': [x]},
-        outputs={'Out': [tmp]},
+        outputs={'Out': [out]},
         attrs={'step': value})
-    return tmp
+    return out
 
 
 def array_write(x, i, array=None, main_program=None):
@@ -865,3 +937,16 @@ def array_read(array, i, main_program=None):
                 'I': [i]},
         outputs={'Out': [out]})
     return out
+
+
+def shrink_memory(x, i, table, main_program=None):
+    helper = LayerHelper('shrink_memory', **locals())
+    out = helper.create_tmp_variable(dtype=x.data_type)
+    helper.append_op(
+        type='shrink_rnn_memory',
+        inputs={'X': [x],
+                'I': [i],
+                'RankTable': [table]},
+        outputs={'Out': [out]},
+        attrs={})
+    return out

python/paddle/v2/framework/tests/test_accuracy_op.py

@@ -26,5 +26,4 @@ class TestAccuracyOp(OpTest):
 
 
 if __name__ == '__main__':
-    exit(0)
     unittest.main()

python/paddle/v2/framework/tests/test_clip_by_norm_op.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestClipByNormOp(OpTest):
+    def setUp(self):
+        self.max_relative_error = 0.006
+        self.initTestCase()
+        input = np.random.random(self.shape).astype("float32")
+        input[np.abs(input) < self.max_relative_error] = 0.5
+        self.op_type = "clip_by_norm"
+        self.inputs = {'X': input, }
+        self.attrs = {}
+        self.attrs['max_norm'] = self.max_norm
+        norm = np.sqrt(np.sum(np.square(input)))
+        if norm > self.max_norm:
+            output = self.max_norm * input / norm
+        else:
+            output = input
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def initTestCase(self):
+        self.shape = (100, )
+        self.max_norm = 1.0
+
+
+class TestCase1(TestClipByNormOp):
+    def initTestCase(self):
+        self.shape = (100, )
+        self.max_norm = 1e20
+
+
+class TestCase2(TestClipByNormOp):
+    def initTestCase(self):
+        self.shape = (16, 16)
+        self.max_norm = 0.1
+
+
+class TestCase3(TestClipByNormOp):
+    def initTestCase(self):
+        self.shape = (4, 8, 16)
+        self.max_norm = 1.0
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_lod_tensor_array_ops.py

@@ -4,6 +4,7 @@ import numpy
 import paddle.v2.framework.layers as layers
 from paddle.v2.framework.framework import Program
 from paddle.v2.framework.executor import Executor
+from paddle.v2.framework.backward import append_backward_ops
 
 
 class TestCPULoDTensorArrayOps(unittest.TestCase):
@@ -123,5 +124,42 @@ class TestCPULoDTensorArrayOps(unittest.TestCase):
         self.assertEqual(actual.lod(), expect.lod())
 
 
+class TestCPULoDTensorArrayOpGrad(unittest.TestCase):
+    def test_grad(self):
+        place = core.CPUPlace()
+        program = Program()
+
+        x = layers.data(
+            name='x',
+            shape=[1],
+            data_type='float32',
+            main_program=program,
+            stop_gradient=False)
+        table = layers.lod_rank_table(x, level=0, main_program=program)
+        array = layers.lod_tensor_to_array(x, table, main_program=program)
+        result = layers.array_to_lod_tensor(array, table, main_program=program)
+
+        mean = layers.mean(x=result, main_program=program)
+
+        append_backward_ops(mean)
+
+        tensor = core.LoDTensor()
+        tensor.set(numpy.arange(10).reshape(10, 1).astype('float32'), place)
+        tensor.set_lod([[0, 3, 9, 10]])
+
+        g_vars = program.global_block().var(x.name + "@GRAD")
+
+        exe = Executor(place)
+        g_out = [
+            item.sum()
+            for item in map(
+                numpy.array,
+                exe.run(program, feed={'x': tensor}, fetch_list=[g_vars]))
+        ]
+        g_out_sum = numpy.array(g_out).sum()
+
+        self.assertAlmostEqual(1.0, g_out_sum, delta=0.1)
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/framework/tests/test_lstm_op.py

@@ -117,8 +117,9 @@ class TestLstmOp(OpTest):
         self.act_cell = 'tanh'
         self.act_cand = 'tanh'
 
-        self.has_initial_state = True
+        self.has_initial_state = False
         self.is_reverse = False
+        self.use_peepholes = True
 
     def setUp(self):
         self.set_argument()
@@ -128,18 +129,28 @@ class TestLstmOp(OpTest):
         N = len(self.lod[0]) - 1
 
         x = np.random.normal(size=(T, 4 * self.D)).astype('float64')
-        h0 = np.zeros((N, self.D)).astype('float64')
-        c0 = np.zeros((N, self.D)).astype('float64')
+        if self.has_initial_state:
+            h0 = np.random.normal(size=(N, self.D)).astype('float64')
+            c0 = np.random.normal(size=(N, self.D)).astype('float64')
+        else:
+            h0 = np.zeros((N, self.D)).astype('float64')
+            c0 = np.zeros((N, self.D)).astype('float64')
         w = np.random.normal(size=(self.D, 4 * self.D)).astype('float64')
-        b = np.random.normal(size=(1, 7 * self.D)).astype('float64')
+        if self.use_peepholes:
+            b = np.random.normal(size=(1, 7 * self.D)).astype('float64')
+        else:
+            b = np.random.normal(size=(1, 4 * self.D)).astype('float64')
 
         w_b = b[:, 0:4 * self.D]
-        w_c = b[:, 4 * self.D:]
+        w_c = b[:, 4 * self.D:] if self.use_peepholes else None
         h, c = lstm(x, self.lod, h0, c0, w, w_b, w_c, self.is_reverse,
                     ACTVATION[self.act_gate], ACTVATION[self.act_cell],
                     ACTVATION[self.act_cand])
 
-        self.inputs = {'Input': (x, self.lod), 'Weight': w, 'Bias': b}
+        self.inputs = {'Input': (x, self.lod), 'Weight': w}
+
+        self.inputs['Bias'] = b
+
         if self.has_initial_state:
             self.inputs['H0'] = h0
             self.inputs['C0'] = c0
@@ -149,17 +160,16 @@ class TestLstmOp(OpTest):
             'Cell': (c, self.lod),
         }
         self.attrs = {
-            'usePeepholes': True,
-            'isReverse': self.is_reverse,
-            'gateActivation': self.act_gate,
-            'cellActivation': self.act_cell,
-            'candidateActivation': self.act_cand
+            'use_peepholes': self.use_peepholes,
+            'is_reverse': self.is_reverse,
+            'gate_activation': self.act_gate,
+            'cell_activation': self.act_cell,
+            'candidate_activation': self.act_cand
         }
 
     def test_check_output(self):
         self.check_output(atol=1e-8)
 
-    #TODO(qingqing) add more unit testing case
     def test_check_grad(self):
         # TODO(qingqing) remove folowing lines after the check_grad is refined.
         N = len(self.lod[0]) - 1
@@ -170,7 +180,7 @@ class TestLstmOp(OpTest):
             ['Input', 'Weight', 'Bias'], ['Hidden'], max_relative_error=5e-4)
 
 
-class TestLstmOpHasNoInitial(TestLstmOp):
+class TestLstmOpHasInitial(TestLstmOp):
     def set_argument(self):
         self.lod = [[0, 2, 5, 7]]
         self.D = 16
@@ -179,8 +189,69 @@ class TestLstmOpHasNoInitial(TestLstmOp):
         self.act_cell = 'tanh'
         self.act_cand = 'tanh'
 
-        self.has_initial_state = False
+        self.has_initial_state = True
         self.is_reverse = True
+        self.use_peepholes = True
+
+    def test_check_grad(self):
+        # TODO(qingqing) remove folowing lines after the check_grad is refined.
+        N = len(self.lod[0]) - 1
+        self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
+        self.outputs['BatchCellPreAct'] = np.zeros(
+            (N, self.D)).astype('float64')
+        self.check_grad(
+            ['Input', 'Weight', 'Bias', 'H0', 'C0'], ['Hidden'],
+            max_relative_error=5e-4)
+
+    def test_check_grad_ingore_bias(self):
+        N = len(self.lod[0]) - 1
+        self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
+        self.outputs['BatchCellPreAct'] = np.zeros(
+            (N, self.D)).astype('float64')
+        self.check_grad(
+            ['Input', 'Weight'], ['Hidden'],
+            max_relative_error=5e-4,
+            no_grad_set=set('Bias'))
+
+    def test_check_grad_ingore_weight(self):
+        N = len(self.lod[0]) - 1
+        self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
+        self.outputs['BatchCellPreAct'] = np.zeros(
+            (N, self.D)).astype('float64')
+        self.check_grad(
+            ['Input', 'Bias'], ['Hidden'],
+            max_relative_error=5e-4,
+            no_grad_set=set('Weight'))
+
+    def test_check_grad_ingore_input(self):
+        N = len(self.lod[0]) - 1
+        self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
+        self.outputs['BatchCellPreAct'] = np.zeros(
+            (N, self.D)).astype('float64')
+        self.check_grad(
+            ['Weight', 'Bias'], ['Hidden'],
+            max_relative_error=5e-4,
+            no_grad_set=set('Input'))
+
+    def test_check_grad_ingore_h0(self):
+        N = len(self.lod[0]) - 1
+        self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
+        self.outputs['BatchCellPreAct'] = np.zeros(
+            (N, self.D)).astype('float64')
+        self.check_grad(
+            ['Input', 'Weight', 'Bias', 'C0'], ['Hidden'],
+            max_relative_error=5e-4,
+            no_grad_set=set('H0'))
+
+    def test_check_grad_ingore_c0(self):
+        N = len(self.lod[0]) - 1
+        self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
+        self.outputs['BatchCellPreAct'] = np.zeros(
+            (N, self.D)).astype('float64')
+        self.check_grad(
+            ['Input', 'Weight', 'Bias', 'H0'], ['Hidden'],
+            max_relative_error=5e-4,
+            no_grad_set=set('C0'))
 
 
 class TestLstmOpRerverse(TestLstmOp):
@@ -192,8 +263,23 @@ class TestLstmOpRerverse(TestLstmOp):
         self.act_cell = 'tanh'
         self.act_cand = 'tanh'
 
-        self.has_initial_state = True
+        self.has_initial_state = False
+        self.is_reverse = True
+        self.use_peepholes = True
+
+
+class TestLstmOpNotUsePeepholes(TestLstmOp):
+    def set_argument(self):
+        self.lod = [[0, 2, 5, 7]]
+        self.D = 16
+
+        self.act_gate = 'sigmoid'
+        self.act_cell = 'tanh'
+        self.act_cand = 'tanh'
+
+        self.has_initial_state = False
         self.is_reverse = True
+        self.use_peepholes = False
 
 
 if __name__ == '__main__':

python/paddle/v2/framework/tests/test_pool2d_op.py

@@ -61,8 +61,8 @@ class TestPool2d_Op(OpTest):
             'strides': self.strides,
             'paddings': self.paddings,
             'ksize': self.ksize,
-            'poolingType': self.pool_type,
-            'globalPooling': self.global_pool,
+            'pooling_type': self.pool_type,
+            'global_pooling': self.global_pool,
         }
 
         self.outputs = {'Out': output.astype('float32')}

python/paddle/v2/framework/tests/test_pool3d_op.py

@@ -67,8 +67,8 @@ class TestPool3d_Op(OpTest):
             'strides': self.strides,
             'paddings': self.paddings,
             'ksize': self.ksize,
-            'poolingType': self.pool_type,
-            'globalPooling': self.global_pool,
+            'pooling_type': self.pool_type,
+            'global_pooling': self.global_pool,
         }
 
         self.outputs = {'Out': output.astype('float32')}

python/paddle/v2/framework/tests/test_pool_max_op.py

@@ -86,7 +86,7 @@ class TestMaxPoolWithIndex_Op(OpTest):
             'strides': self.strides,
             'paddings': self.paddings,
             'ksize': self.ksize,
-            'globalPooling': self.global_pool,
+            'global_pooling': self.global_pool,
         }
 
         self.inputs = {'X': input}

python/paddle/v2/framework/tests/test_shrink_rnn_memory.py

+import unittest
+import paddle.v2.framework.core as core
+from paddle.v2.framework.executor import Executor
+import paddle.v2.framework.layers as layers
+from paddle.v2.framework.backward import append_backward_ops
+from paddle.v2.framework.framework import g_main_program
+import numpy
+
+
+class TestShrinkRNNMemory(unittest.TestCase):
+    def test_shrink_rnn_memory(self):
+        x = layers.data('x', shape=[100], data_type='float32')
+        x.stop_gradient = False
+        table = layers.lod_rank_table(x=x)
+        i = layers.zeros(dtype='int64', shape=[1])
+        mem1 = layers.shrink_memory(x=x, i=i, table=table)
+        i = layers.increment(x=i)
+        i.stop_gradient = True
+        mem2 = layers.shrink_memory(x=mem1, i=i, table=table)
+        i = layers.increment(x=i)
+        i.stop_gradient = True
+        mem3 = layers.shrink_memory(x=mem2, i=i, table=table)
+
+        cpu = core.CPUPlace()
+        tensor = core.LoDTensor()
+        tensor.set_lod([[0, 2, 5, 6]])
+        tensor_np = numpy.random.random(size=(3, 100)).astype('float32')
+        tensor.set(tensor_np, cpu)
+        exe = Executor(cpu)
+        outs = map(numpy.array,
+                   exe.run(feed={'x': tensor}, fetch_list=[mem1, mem2, mem3]))
+        self.assertTrue(numpy.allclose(tensor_np[0:3], outs[0]))
+        self.assertTrue(numpy.allclose(tensor_np[0:2], outs[1]))
+        self.assertTrue(numpy.allclose(tensor_np[0:1], outs[2]))
+
+        mem3_mean = layers.mean(x=mem3)
+        append_backward_ops(loss=mem3_mean)
+        x_grad = map(numpy.array,
+                     exe.run(feed={'x': tensor},
+                             fetch_list=[
+                                 g_main_program.global_block().var('x@GRAD')
+                             ]))[0]
+        self.assertAlmostEqual(1.0, x_grad.sum(), delta=0.1)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_understand_sentiment_lstm.py

+import paddle.v2 as paddle
+import paddle.v2.framework.layers as layers
+import paddle.v2.framework.core as core
+import paddle.v2.framework.optimizer as optimizer
+
+from paddle.v2.framework.framework import g_main_program, g_startup_program
+from paddle.v2.framework.executor import Executor
+
+import numpy as np
+
+
+def lstm_net(dict_dim, class_dim=2, emb_dim=32, seq_len=80, batch_size=50):
+    data = layers.data(
+        name="words",
+        shape=[seq_len * batch_size, 1],
+        append_batch_size=False,
+        data_type="int64")
+    label = layers.data(
+        name="label",
+        shape=[batch_size, 1],
+        append_batch_size=False,
+        data_type="int64")
+
+    emb = layers.embedding(input=data, size=[dict_dim, emb_dim])
+    emb = layers.reshape(x=emb, shape=[batch_size, seq_len, emb_dim])
+    emb = layers.transpose(x=emb, axis=[1, 0, 2])
+
+    c_pre_init = layers.fill_constant(
+        dtype=emb.data_type, shape=[batch_size, emb_dim], value=0.0)
+    layer_1_out = layers.lstm(emb, c_pre_init=c_pre_init, hidden_dim=emb_dim)
+    layer_1_out = layers.transpose(x=layer_1_out, axis=[1, 0, 2])
+
+    prediction = layers.fc(input=layer_1_out, size=class_dim, act="softmax")
+    cost = layers.cross_entropy(input=prediction, label=label)
+
+    avg_cost = layers.mean(x=cost)
+    adam_optimizer = optimizer.AdamOptimizer(learning_rate=0.002)
+    opts = adam_optimizer.minimize(avg_cost)
+    acc = layers.accuracy(input=prediction, label=label)
+
+    return avg_cost, acc
+
+
+def to_lodtensor(data, place):
+    seq_lens = [len(seq) for seq in data]
+    cur_len = 0
+    lod = [cur_len]
+    for l in seq_lens:
+        cur_len += l
+        lod.append(cur_len)
+    flattened_data = np.concatenate(data, axis=0).astype("int64")
+    flattened_data = flattened_data.reshape([len(flattened_data), 1])
+    res = core.LoDTensor()
+    res.set(flattened_data, place)
+    res.set_lod([lod])
+    return res
+
+
+def chop_data(data, chop_len=80, batch_len=50):
+    data = [(x[0][:chop_len], x[1]) for x in data if len(x[0]) >= chop_len]
+
+    return data[:batch_len]
+
+
+def prepare_feed_data(data, place):
+    tensor_words = to_lodtensor(map(lambda x: x[0], data), place)
+
+    label = np.array(map(lambda x: x[1], data)).astype("int64")
+    label = label.reshape([50, 1])
+    tensor_label = core.LoDTensor()
+    tensor_label.set(label, place)
+
+    return tensor_words, tensor_label
+
+
+def main():
+    word_dict = paddle.dataset.imdb.word_dict()
+    cost, acc = lstm_net(dict_dim=len(word_dict), class_dim=2)
+
+    batch_size = 100
+    train_data = paddle.batch(
+        paddle.reader.buffered(
+            paddle.dataset.imdb.train(word_dict), size=batch_size * 10),
+        batch_size=batch_size)
+
+    data = chop_data(next(train_data()))
+
+    place = core.CPUPlace()
+    tensor_words, tensor_label = prepare_feed_data(data, place)
+    exe = Executor(place)
+    exe.run(g_startup_program)
+
+    while True:
+        outs = exe.run(g_main_program,
+                       feed={"words": tensor_words,
+                             "label": tensor_label},
+                       fetch_list=[cost, acc])
+        cost_val = np.array(outs[0])
+        acc_val = np.array(outs[1])
+
+        print("cost=" + str(cost_val) + " acc=" + str(acc_val))
+        if acc_val > 0.9:
+            break
+
+
+if __name__ == '__main__':
+    main()

python/paddle/v2/image.py

-import numpy as np
-try:
-    import cv2
-except ImportError:
-    cv2 = None
-import os
-import tarfile
-import cPickle
-
-__all__ = [
-    "load_image_bytes", "load_image", "resize_short", "to_chw", "center_crop",
-    "random_crop", "left_right_flip", "simple_transform", "load_and_transform",
-    "batch_images_from_tar"
-]
 """
 This file contains some common interfaces for image preprocess.
 Many users are confused about the image layout. We introduce
 the image layout as follows.
 
 - CHW Layout
+
   - The abbreviations: C=channel, H=Height, W=Width
   - The default layout of image opened by cv2 or PIL is HWC.
     PaddlePaddle only supports the CHW layout. And CHW is simply
     a transpose of HWC. It must transpose the input image.
 
 - Color format: RGB or BGR
+
   OpenCV use BGR color format. PIL use RGB color format. Both
   formats can be used for training. Noted that, the format should
   be keep consistent between the training and inference peroid.
 """
+import numpy as np
+try:
+    import cv2
+except ImportError:
+    cv2 = None
+import os
+import tarfile
+import cPickle
+
+__all__ = [
+    "load_image_bytes", "load_image", "resize_short", "to_chw", "center_crop",
+    "random_crop", "left_right_flip", "simple_transform", "load_and_transform",
+    "batch_images_from_tar"
+]
 
 
 def batch_images_from_tar(data_file,
@@ -36,17 +38,18 @@ def batch_images_from_tar(data_file,
                           num_per_batch=1024):
     """
     Read images from tar file and batch them into batch file.
-    param data_file: path of image tar file
-    type data_file: string
-    param dataset_name: 'train','test' or 'valid'
-    type dataset_name: string
-    param img2label: a dic with image file name as key 
+
+    :param data_file: path of image tar file
+    :type data_file: string
+    :param dataset_name: 'train','test' or 'valid'
+    :type dataset_name: string
+    :param img2label: a dic with image file name as key 
                     and image's label as value
-    type img2label: dic
-    param num_per_batch: image number per batch file
-    type num_per_batch: int
-    return: path of list file containing paths of batch file
-    rtype: string
+    :type img2label: dic
+    :param num_per_batch: image number per batch file
+    :type num_per_batch: int
+    :return: path of list file containing paths of batch file
+    :rtype: string
     """
     batch_dir = data_file + "_batch"
     out_path = "%s/%s" % (batch_dir, dataset_name)
@@ -99,14 +102,16 @@ def load_image_bytes(bytes, is_color=True):
     Example usage:
     
     .. code-block:: python
+
         with open('cat.jpg') as f:
             im = load_image_bytes(f.read())
 
     :param bytes: the input image bytes array.
-    :type file: str
+    :type bytes: str
     :param is_color: If set is_color True, it will load and
                      return a color image. Otherwise, it will
                      load and return a gray image.
+    :type is_color: bool
     """
     flag = 1 if is_color else 0
     file_bytes = np.asarray(bytearray(bytes), dtype=np.uint8)
@@ -121,6 +126,7 @@ def load_image(file, is_color=True):
     Example usage:
     
     .. code-block:: python
+
         im = load_image('cat.jpg')
 
     :param file: the input image path.
@@ -128,6 +134,7 @@ def load_image(file, is_color=True):
     :param is_color: If set is_color True, it will load and
                      return a color image. Otherwise, it will
                      load and return a gray image.
+    :type is_color: bool
     """
     # cv2.IMAGE_COLOR for OpenCV3
     # cv2.CV_LOAD_IMAGE_COLOR for older OpenCV Version
@@ -147,6 +154,7 @@ def resize_short(im, size):
     Example usage:
     
     .. code-block:: python
+
         im = load_image('cat.jpg')
         im = resize_short(im, 256)
     
@@ -175,6 +183,7 @@ def to_chw(im, order=(2, 0, 1)):
     Example usage:
     
     .. code-block:: python
+
         im = load_image('cat.jpg')
         im = resize_short(im, 256)
         im = to_chw(im)
@@ -196,6 +205,7 @@ def center_crop(im, size, is_color=True):
     Example usage:
     
     .. code-block:: python
+
         im = center_crop(im, 224)
     
     :param im: the input image with HWC layout.
@@ -223,6 +233,7 @@ def random_crop(im, size, is_color=True):
     Example usage:
     
     .. code-block:: python
+
         im = random_crop(im, 224)
     
     :param im: the input image with HWC layout.
@@ -251,6 +262,7 @@ def left_right_flip(im):
     Example usage:
     
     .. code-block:: python
+
         im = left_right_flip(im)
     
     :paam im: input image with HWC layout
@@ -275,6 +287,7 @@ def simple_transform(im,
     Example usage:
     
     .. code-block:: python
+
         im = simple_transform(im, 256, 224, True)
 
     :param im: The input image with HWC layout.
@@ -285,6 +298,11 @@ def simple_transform(im,
     :type crop_size: int
     :param is_train: Whether it is training or not.
     :type is_train: bool
+    :param is_color: whether the image is color or not.
+    :type is_color: bool
+    :param mean: the mean values, which can be element-wise mean values or 
+                 mean values per channel.
+    :type mean: numpy array | list
     """
     im = resize_short(im, resize_size)
     if is_train:
@@ -324,6 +342,7 @@ def load_and_transform(filename,
     Example usage:
     
     .. code-block:: python
+
         im = load_and_transform('cat.jpg', 256, 224, True)
 
     :param filename: The file name of input image.
@@ -334,6 +353,11 @@ def load_and_transform(filename,
     :type crop_size: int
     :param is_train: Whether it is training or not.
     :type is_train: bool
+    :param is_color: whether the image is color or not.
+    :type is_color: bool
+    :param mean: the mean values, which can be element-wise mean values or 
+                 mean values per channel.
+    :type mean: numpy array | list
     """
     im = load_image(filename)
     im = simple_transform(im, resize_size, crop_size, is_train, is_color, mean)

python/paddle/v2/optimizer.py

@@ -102,7 +102,7 @@ class Momentum(Optimizer):
 
     ..  math::
 
-        v_{t} &= k * v_{t-1} - \\gamma_t / (g_{t} + \\lambda w_{t-1}) \\\\
+        v_{t} &= k * v_{t-1} - \\gamma_t (g_{t} + \\lambda w_{t-1}) \\\\
         w_{t} &= w_{t-1} + v_{t} \\\\
 
     where, :math:`k` is momentum, :math:`\\lambda` is decay rate,